Flux Approximation Schemes For Flow Problems Using Local Boundary Value Problems

Flux approximation schemes for flow problems using
local boundary value problems

Kumar, N.
Published: 20/12/2017
Document Version
Publisher’s PDF, also known as Version of Record (includes final page, issue and volume numbers)
Please check the document version of this publication:
• A submitted manuscript is the author's version of the article upon submission and before peer-review. There can be important differences
between the submitted version and the official published version of record. People interested in the research are advised to contact the
author for the final version of the publication, or visit the DOI to the publisher's website.
• The final author version and the galley proof are versions of the publication after peer review.
• The final published version features the final layout of the paper including the volume, issue and page numbers.
Link to publication
Citation for published version (APA):

Kumar, N. (2017). Flux approximation schemes for flow problems using local boundary value problems
Eindhoven: Technische Universiteit Eindhoven
General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners
and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.
• Users may download and print one copy of any publication from the public portal for the purpose of private study or research.
• You may not further distribute the material or use it for any profit-making activity or commercial gain
• You may freely distribute the URL identifying the publication in the public portal ?
Take down policy
If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately
and investigate your claim.
Download date: 01. Jan. 2019

Flux Approximation Schemes for
Flow Problems using
Local Boundary Value Problems
Nikhil Kumar
This work is part of the Industrial Partnership Programme (IPP) Computa-
tional Sciences for Energy Research (project number: 101760) of the (former)
Foundation for Fundamental Research on Matter (FOM), which is (now) part
of the Netherlands Organization for Scientific Research (NWO). This research
programme is co-financed by Shell Global Solutions International B.V.
A catalogue record is available from the Eindhoven University of Technology

Library
ISBN: 978-90-386-4391-5
Copyright © 2017 by N. Kumar.

All rights are reserved. No part of this publication may be reproduced, stored
in a retrieval system, or transmitted, in any form or by any means, electronic,
mechanical, photocopying, recording or otherwise, without prior permission of
the author.
Flux Approximation Schemes for Flow Problems using
Local Boundary Value Problems
PROEFSCHRIFT
ter verkrijging van de graad van doctor aan de

Technische Universiteit Eindhoven, op gezag van de
rector magnificus prof.dr.ir. F.P.T. Baaijens, voor een
commissie aangewezen door het College voor
Promoties, in het openbaar te verdedigen
op woensdag 20 december 2017 om 14:00 uur
door
Nikhil Kumar
geboren te Jaipur, India

Dit proefschrift is goedgekeurd door de promotoren en de samen-
stelling van de promotiecommissie is als volgt:
voorzitter: prof.dr. M.A. Peletier

promotor: prof.dr.ir. B. Koren
co-promotor: dr.ir. J.H.M. ten Thije Boonkkamp
leden: prof.dr.ir. C. Vuik (TU Delft)
prof.dr. A.E.P Veldman (RUG)
prof.dr.ir. E.H. van Brummelen
adviseurs dr.ir. B. Sanderse (CWI)
dr. A. Linke (WIAS, Berlijn)
Het onderzoek dat in dit proefschrift wordt beschreven is uitgevoerd

in overeenstemming met de TU/e Gedragscode Wetenschapsbeoe-
fening.
Contents
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Spatial discretization . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Main results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Outline of the thesis . . . . . . . . . . . . . . . . . . . . . . . . 6
2 Discretization techniques 9
2.1 Convection-diffusion equation . . . . . . . . . . . . . . . . . . . 10
2.2 Incompressible Navier-Stokes equations . . . . . . . . . . . . . 16
3 Convection-diffusion equation 21
3.1 Convection-diffusion equation . . . . . . . . . . . . . . . . . . . 22
3.2 Inhomogeneous local BVP approximation . . . . . . . . . . . . 28
3.3 Numerical dissipation and dispersion . . . . . . . . . . . . . . . 32
3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4 Convective discretization of the Navier-Stokes equations 37

4.1 Convective discretization . . . . . . . . . . . . . . . . . . . . . . 38
4.2 Integral representation of cell-face
velocities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.3 Closure of the model . . . . . . . . . . . . . . . . . . . . . . . . 45
4.4 Iterative computation of the cell-face
velocities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.5 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5 Complete-flux schemes 57
5.1 Integral representation of the flux . . . . . . . . . . . . . . . . . 58
5.2 Complete-flux scheme for the viscous
Burgers equation . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.3 Nonlinear local BVP approximation . . . . . . . . . . . . . . . 69
i
5.4 Flux in the inviscid limit . . . . . . . . . . . . . . . . . . . . . . 72
5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6 Complete-flux scheme for the Navier-Stokes equations 77

6.1 Finite volume method . . . . . . . . . . . . . . . . . . . . . . . 78
6.2 Integral representation of the fluxes . . . . . . . . . . . . . . . . 79
6.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
7 Nonlinear flux approximation 87

7.1 Flux from local two-point BVP . . . . . . . . . . . . . . . . . . 88
7.2 Numerical flux function . . . . . . . . . . . . . . . . . . . . . . 89
7.3 Existence of numerical flux . . . . . . . . . . . . . . . . . . . . 92
7.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
8 Conclusion 105
8.1 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . 105
8.2 Recommendations for future work . . . . . . . . . . . . . . . . 107
Bibliography 109
Summary 118
Curriculum Vitae 121
List of publications 123
Acknowledgments 125
ii
Abbreviations
BVP boundary value problem
CFS complete-flux scheme
DAE differential-algebraic equation
FVM finite-volume method
HFS homogeneous flux scheme
ODE ordinary differential equation
PDE partial differential equation
TVD total variation diminishing
TVNI total variation non-increasing
iii
List of symbols
ν diffusion coefficient
ν num numerical diffusion coefficient
art
ν artificial diffusion coefficient
Re Reynolds number
P grid Péclet number
σ scaled coordinate
L differential operator for the steady convection-diffusion equation
Lh discrete form of the differential operator L
TV total variation
Ω mass matrix associated with the spatial discretization
M discrete divergence operator
N nonlinear convection operator with components (N u , N v )
G discrete gradient operator with components (Gu , Gv )
D discrete diffusion operator with components (Du , Dv )
L discrete Laplacian operator
iv
List of functions
1 + z
Artanh(z) = log
1−z
tanh(z)
tanhc(z) =
z
z
B(z) =
ez − 1
ez/2 − 1 − z/2
C(z) =
z(ez − 1)
D(z) = z tanh(z)
1 2
E(z) = 2tanhc(z) + tanh(z)
4
1
W (z) =
ez +1
ez − 1 − z
W1 (z) =
z(ez − 1)
eσz − 1
H1 (σ; z) =
ez − 1
ez − 1 − z
H2 (z) =
z 2 (ez + 1)
v
Chapter 1
Introduction
The conservation laws governing fluid flow involve a combination of convective

and diffusive processes. Mathematically, these conservation laws are given by
a system of partial differential equations (PDEs). Apart from a few special
cases, analytical solutions to these PDEs are not available or quite difficult
to compute. Thus, we have to rely on numerical solutions of the governing
equations, for a better understanding of the dynamics of the flow. The modeling
of the interaction between convection and diffusion is an arduous task in the
numerical analysis of PDEs. This is mainly because of the striking difference
in the two limiting cases: for a purely convection-driven flow, the resulting
PDE is hyperbolic, whereas for diffusion-driven flow, it is parabolic. In the
flow problems considered in this thesis, the interplay between convection and
diffusion is governed by the diffusion coefficient (the Reynolds number in case of
the Navier-Stokes equations). For smaller values of the diffusion coefficient, the
PDE becomes singularly perturbed, with the solution exhibiting thin boundary
layers. A detailed explanation of PDEs and singularly perturbed problems can
be found in [59, 71, 96]. Designing a numerical scheme which gives an accurate
approximation in all regimes is a challenging task.
The computation of numerical solutions can be broadly categorized into,
the discretization of the governing PDEs, resulting in a system of algebraic
equations and solving the algebraic system of equations [38, 62]. For unsteady
flow problems, the discretization process is split into the spatial and the tem-
poral part. The spatial discretization replaces the spatial continuum by a finite
number of points or cells in which the numerical values of the variables are de-
termined. For the spatial discretization of PDEs, there are a number of choices,
the finite difference, finite volume, finite element and spectral methods being
the most commonly used techniques. The temporal discretization or time in-
tegration involves approximation of the time derivative and gives an algebraic
system of equations at each time step [53].
1
Introduction
In this thesis, we focus mainly on the spatial discretization using the finite-
volume method for the viscous Burgers equation [11] and the incompressible
Navier-Stokes equations. The spatial discretization schemes discussed in the
thesis can be applied to other models as well, involving the convection-diffusion
phenomenon. In the following, we mention a couple of these models:
• Drift-diffusion model. The drift-diffusion model is used for describing

the behavior of semiconductor devices. The model is given by a Poisson
equation for the electrostatic potential and a continuity equation for holes
and electrons [28, 81]. Simulating semiconductor behavior using the drift-
diffusion model has led to the development of several numerical methods
[64], like the Scharfetter-Gummel scheme and its extension [80, 89]. Apart
from semiconductors, the drift-diffusion model is also extensively used in
the modeling of plasmas, using the multi-fluid model, which treats each
species in a plasma as a fluid [32, 43, 67].
• Black-Scholes model. The Black-Scholes model, first published by Fisher

Black and Myron Scholes [7], describes the price of an option over time
in a financial market. The model provided mathematical legitimacy to
option trading in the stock markets across the world [56]. Mathematically,
the Black-Scholes model is similar to the unsteady convection-diffusion-
reaction equation. Due to the prevalence of the model in the financial
markets, it is essential to have robust numerical schemes for the model.
1.1 Motivation
The research described in this thesis has been carried out within the framework
of the topic ‘Wind op Zee’ (Offshore wind) of the Dutch national ‘Topsector
Energie’. In essence, the goal of the research in this thesis is to further pave
the way for robust, accurate and efficient computations of turbulent air flows
in offshore wind farms.
Figure 1.1 shows the Horns Rev 2 offshore wind farm under stable atmo-
spheric conditions [41]. The grouping of wind turbines in an offshore wind
farm raises two issues: wake velocity deficits and higher loads on the blades
due to increased turbulence. In order to address both issues it is essential to
understand the wake dynamics of the turbines. The physics of the wind turbine
wake can be described by the incompressible Navier-Stokes equations, with the
atmospheric velocities ranging between 4-25 m/s [30]. Note that for modeling
the aerodynamics at the blade tips, with increasing rotor diameters, in future,
it may become necessary to take into account compressibility effects. To model
the wake it is essential to have robust, accurate and efficient numerical methods
2
Introduction
Figure 1.1: Photograph of the Horns Rev 2 offshore wind farm under stable
atmospheric conditions (wind speed 13 m/s at hub height), taken by Bel Air
Aviation Denmark.
for the incompressible Navier-Stokes equations, given by:
∇ · u = 0, (1.1a)
∂u 1 2
+ ∇ · (uu) = −∇p + ∇ u, (1.1b)
∂t Re
which govern the flow of a Newtonian incompressible fluid with velocity field u,
pressure p and Reynolds number Re. The above system of equations is derived
from the principle of conservation of mass and momentum for the fluid, more
details can be found in [5, 102].
For turbulent flows, the presence of the nonlinear convective term introduces
a wide range of length and time scales depending on the Reynolds number, the
ratio of convective and viscous forces, [4, 16, 47, 83]. For modern-day wind tur-
bines having rotor diameters up to 120 m, with Reynolds numbers of the order
108 , resolving all length scales using direct numerical simulation (DNS) is still
unrealistic. Thus turbulence models are constructed to account for the effects
at small scale. A large number of turbulence models have been constructed in
the past decades, see e.g. [65, 73, 77, 104], with RANS (Reynolds-Averaged
Navier-Stokes) and LES (Large Eddy Simulation) being the most commonly
used models. For modeling the wake of a wind farm LES is preferred due to its
ability to handle unsteady turbulent flows dominated by large scale structures
and turbulent mixing [45, 74].
3
Introduction
When using LES, one should be very careful with the discretization of the
nonlinear convective term. In [1], the authors provide an overview of finite-
volume discretization schemes for LES. The central scheme does not add any
non-physical dissipation to the system, but suffers from spurious oscillations,
particularly at higher Reynolds numbers [36, 105]. The spurious oscillations
can be removed by explicitly adding numerical dissipation for stabilization or
by using an upwind scheme. However, using an upwind scheme can influence
the energy cascades from large to small scales, due to numerical dissipation
[57]. To avoid this, central schemes are sometimes used in combination with
higher-order upwind schemes [6, 84]. As a drawback, higher-order schemes
may significantly increase the computational cost of the method. Not only the
spatial discretization, but also the time integration can introduce numerical
dissipation and introduce instabilities in the numerical scheme. The time in-
tegration using the implicit Runge-Kutta methods as described in [74, 75] is
shown to be stable and energy-conserving. In this thesis, we present spatial
discretization methods, that do not suffer from spurious oscillations and intro-
duce no or very little numerical dissipation, without any significant increase in
the computational cost.
1.2 Spatial discretization

With the aim to derive robust and accurate numerical methods, without signific-
antly increasing the computational cost, we impose the following requirements
on our spatial discretization:
• be completely free of any spurious oscillations for any grid size,
• have a three-point coupling, corresponding to each spatial direction,
• have as little numerical dissipation as possible,
• be uniformly second-order accurate,
• take into account all the forces acting on the fluid locally.
We present spatial discretization methods for flow problems involving non-
linear convection using local boundary value problems (BVPs). The scheme
is inspired by the flux approximation schemes for the convection-diffusion-
reaction equation presented in [87, 90]. The scheme is based on the principle of
solving the conservation law governing the local flow, such that the numerical
approximation is locally exact. Similar schemes have been proposed in [91, 92],
where the flux is approximated using a local BVP assuming constant coeffi-
cients, resulting in the exponentially fitted scheme. A more detailed account
of exponentially fitted schemes, which are hybrid combinations of the central
and upwind scheme, can be found in [18, 20, 44]. An improvement of the expo-
nentially fitted schemes can be found in [21], where the exponential scheme is
4
Introduction
combined with Keller’s box scheme [46] to produce a stable second-order finite-
difference scheme. Exponentially fitted schemes are widely used for simulating
the drift-diffusion equations for semi-conductor devices, where these schemes
are known as the Scharfetter-Gummel schemes [3, 9].
In our approach, we derive an integral representation for the flux using the
local BVP, including the source term, e.g., the reaction term for the convection-
diffusion-reaction equation and the pressure gradient for the Navier-Stokes
equations, etc. For unsteady problems, the time derivative of the unknown
can also be included as a source term, as in [86]. The integral representation
can be expressed as the sum of a homogeneous part, corresponding to the homo-
geneous solution of the local BVP, and an inhomogeneous part, corresponding
to the particular solution (depending on the source term), respectively. The
numerical flux is subsequently determined from the integral representation us-
ing suitable quadrature rules. In the absence of the source term, the scheme
reduces to the exponentially fitted scheme, [55], which is not uniformly second-
order accurate, but is free of any spurious oscillations. The inclusion of the
source term in the local BVP, results in reduced local truncation errors for the
discretization [55], giving a uniformly second-order approximation.
As mentioned earlier, the flux approximation using a homogeneous local
BVP results in the exponentially fitted scheme, where the numerical scheme
has the same exponential solution as the corresponding differential equation.
For the linear convection-diffusion equation, the exponentially fitted scheme
provides an accurate description of the behavior of the local BVP. However, for
nonlinear problems, which may not necessarily possess an exponential solution,
using the homogeneous flux approximation might not result in a locally exact
scheme. For simulating turbulent flows, it is essential that the nonlinearity of
the convective part of the flux is taken into consideration, as linearization may
eliminate the length scales arising due to the nonlinearity of the convective
term. Consequently, we have constructed a nonlinear local BVP scheme for
the approximation of the fluxes in the viscous Burgers equation, where instead
of deriving integral representations of the flux, the numerical flux is computed
directly from a nonlinear equation.
1.3 Main results

The main results in this thesis can be split into two parts, for i) the incom-
pressible Navier-stokes equations, and ii) the viscous Burgers equation.
Incompressible Navier-Stokes equations. The nonlinear convective term

in the momentum balance equations is discretized using local BVPs, with the
discrete pressure gradient and the gradient of the cross-flux as the possible
source terms. Since the convective term is nonlinear the local BVP is linearized
and solved iteratively. The inclusion of the source terms is shown to improve the
5
Introduction
accuracy of the homogeneous approximation, resulting in a uniformly second-

order accurate scheme, which is free of any spurious oscillations.
The complete-flux scheme as presented in [87] has been extended for the
approximation of the fluxes involved in the momentum equations. Again for
the flux approximation, we show that the inclusion of the source terms results
in a uniformly second-order accurate approximation, that is free of any spuri-
ous oscillations.
Viscous Burgers equation. We derive the integral representations using

nonlinear local BVP for the viscous Burgers equation. The integral representa-
tion thus obtained is implicit, i.e, depends on the solution of the nonlinear local
BVP. To simplify the flux approximation, we linearize the local BVP using the
cell-face velocity. For an approximation that is consistent in the inviscid limit,
it is required that the cell-face velocity is computed using either the upwind or
the homogeneous 1-D local BVP scheme. We also show that applying stand-
ard quadrature rules to the implicit integral representation does not provide a
consistent numerical flux.
The nonlinear local BVP scheme for the viscous Burgers equation is pro-
posed in which the numerical flux is computed from a homogeneous nonlinear
local BVP. The numerical local BVP scheme is shown to provide a numerical
flux for all possible cases. In the inviscid limit, it is shown that the numer-
ical flux is always consistent with the Godunov flux for the inviscid Burgers
equation.
1.4 Outline of the thesis

In Chapter 2 we introduce the necessary theory regarding numerical meth-
ods for partial differential equations. We begin by considering the convection-
diffusion equation and present methods for the spatial and temporal discretiza-
tion. Then we describe the spatial discretization of the incompressible Navier-
Stokes equations using a staggered grid [39] configuration for a second-order
accurate method on uniform grids. The resulting semi-discrete system of equa-
tions is then integrated using implicit Runge-Kutta methods as described in
[75].
In Chapter 3, we present the use of the local BVPs for the spatial discret-
ization of the constant coefficient convection-diffusion equation. The chapter
also discusses the classical discretization methods, viz., the central, upwind and
hybrid schemes. We rehearse the well-known results that the central scheme res-
ults in a second-order accurate approximation, but introduces spurious oscilla-
tions, and that, in contrast, the upwind scheme does not introduce spurious os-
cillations, but introduces numerical dissipation resulting in a first-order accur-
ate approximation. The homogeneous local BVP scheme, which is a weighted
average of the upwind and central scheme, suppresses numerical oscillations
6
Introduction
and exhibits first-order accuracy over coarse grids and second-order accuracy
over finer grids. Further, we demonstrate the significance of the source terms.
We show that the discretization of the convective term using inhomogeneous
local BVPs results in a second-order accurate approximation, which is free of
any non-physical oscillations.
Subsequently, in Chapter 4 we extend the use of local BVPs for the discret-
ization of the nonlinear convective terms in the incompressible Navier-Stokes
equations. The discrete convective terms are computed from linearized local
BVPs. As in the case of the linear convection-diffusion equation, it is observed
that the inclusion of the source term, i.e., the pressure gradient and the gradi-
ent of the transverse flux, provides a more accurate approximation and appears
to remove the numerical diffusion introduced by the homogeneous part of the
approximation. In order to account for the nonlinearity of the convective term,
the local BVP is solved iteratively. The scheme is tested for some classical flow
problems, such as, the Taylor-Green vortex and lid-driven cavity problem. Nu-
merical results for the inhomogeneous approximation scheme exhibit uniform
second-order convergence without introducing any non-physical oscillations.
In Chapter 5, we extend the use of local BVPs for the approximation of the
total flux, i.e., the sum of the convective and the diffusive flux. We present the
integral representations for the flux derived from inhomogeneous local BVPs,
as shown in [87]. The scheme is then extended to the viscous Burgers equa-
tion. We present two approaches to deal with the nonlinear local BVPs. In the
first approach, we linearize the local BVP using the cell-face velocity compon-
ents that are themselves computed from local BVPs, as shown in Chapter 4.
The linearized BVP can then be integrated exactly. In the second approach,
we derive the integral representation for the nonlinear problem and then use
straightforward quadrature rules to derive the numerical flux. We show that
applying quadrature rules to the nonlinear integral representation does not
result in a consistent numerical flux.
The flux approximation scheme using the linearized local BVPs for the vis-
cous Burgers equation is extended to the incompressible Navier-Stokes equa-
tions in Chapter 6. We combine the flux approximation scheme with the Harlow
& Welch [39] staggered grid configuration, using the pressure gradient and the
gradient of the transverse flux as source terms. The inclusion of the source
terms in the BVP is shown to result in a uniformly second-order approxima-
tion.
In Chapter 7, we present another approach for computing the numerical flux
for the viscous Burgers equation using a nonlinear local BVP. In this scheme,
the numerical flux is computed by solving a nonlinear algebraic equation that
is in turn derived from the nonlinear local BVP. The scheme is shown to be
consistent with Godunov’s approximation in the inviscid limit.
Finally, in Chapter 8 we summarize our findings, draw conclusions and
present recommendations for future research.
7
Chapter 2
Discretization techniques
In this chapter, we present an overview of the spatial and temporal discretiz-

ation techniques used in this thesis. To give the details of the spatial discret-
ization and the time integration, we consider the convection-diffusion equation
and the incompressible Navier-Stokes equations. The spatial discretization
for both equations is done using finite-volume methods on uniform Cartesian
grids. Finite-volume methods use the integral form of the conservation law
and therefore involve cell-averaged quantities instead of point values. The in-
tegral formulation is very well suited for problems that contain discontinuities
and results in a conservative discretization. Further, the discretization can be
easily extended to unstructured grids.
For unsteady problems, the spatial discretization results in a system of
ODEs, which is then numerically integrated, also termed time integration or
temporal discretization. For the convection-diffusion equation we may use the
explicit fourth-order Runge-Kutta method for time integration, for which the
discretization errors are negligible. However, explicit methods impose a severe
restriction on the time step size. This can be avoided by using semi-implicit or
implicit methods for time integration, which generally have a larger stability
region and therefore, allow for larger time steps.
The discretization of the incompressible Navier-Stokes equations is more
challenging. In the incompressible Navier-Stokes equations, the pressure acts
as a Lagrange multiplier and ensures that the velocity field is divergence-free
[34, 79]. This coupling of pressure and velocity is described by the pressure-
Poisson equation, which is derived by taking the divergence of the momentum
equations and using the continuity equation. It is essential that the spatial
discretization also exhibits the coupling of the discrete pressure and velocity.
For the spatial discretization of the incompressible Navier-Stokes equations
we use Harlow and Welch’s staggered grid method [39, 99]. Apart from the
coupling of the pressure and velocity, the staggering of variables results in the
9
discrete conservation of mass, momentum and energy (in the inviscid case)
on Cartesian grids [60]. We discuss the second-order accurate discretization
method in two dimensions. The discretization can be extended to higher-order
methods and can also be used for three-dimensional flow problems.
Applying the spatial discretization results in a differential-algebraic equa-
tion (DAE) system, which is integrated in time using Runge-Kutta methods.
Details regarding the application of Runge-Kuttta methods to DAE systems
can be found in [37, 58]. We use the implicit Runge-Kutta methods presented
in [74, 75, 76], which are known to be energy conserving and have non-linear
stability for any given time step size.
This chapter is organized as follows: in Section 2.1 we describe the spatial
and temporal discretization for the convection-diffusion equation. Section 2.2
gives the spatial discretization of the incompressible Navier-Stokes equations
and the time integration of the resulting semi-discrete system.
2.1 Convection-diffusion equation

In the finite-volume method, the computational domain is covered with a fi-
nite number of disjoint control volumes or cells and the integral form of the
conservation law is then discretized over these control volumes. Subsequently,
numerical integration is used to express the integrals in terms of average values,
also called cell-averages [23, 54]. To demonstrate the discretization we consider
the time-dependent convection-diffusion equation
∂φ
+ ∇ · f = s, f = aφ − ν∇φ, (2.1)
∂t
for the unknown φ, where f is the flux function, a the velocity field, ν the
constant diffusion coefficient, and s = s(x, t) the source term. Integrating the
above conservation law on a fixed domain Ω with boundary Γ and using the
divergence theorem gives the integral form of the conservation law, i.e.,
Z I Z
d
φ dV + f · n dS = s dV, (2.2)
dt Ω Γ Ω
where n is the outward unit normal on the boundary Γ. We adopt the cell-
centered approach [100], where the variable φ is approximated at the center of
the cell. For a two-dimensional cell Ωj , j = (i, j) (see Figure 2.1) the conser-
vation law is given by
Z X Z Z
d
φ dA + f · n ds = s dA, (2.3)
dt Ωj Γj,k Ωj
k∈N (j)
where N (j) is the index set of neighboring grid points of xj and Γj,k is the
segment of the boundary Γj connecting the adjacent cells Ωj and Ωk . Approx-
10
j+1
Ωj
j−1
i−1 i i+1
Figure 2.1: Grid structure for the spatial discretization of the convection-
diffusion equation.
imating all integrals in equation (2.3) using the midpoint rule [8], we obtain
the semi-discrete conservation law,
X
φ̇j (t) |Ωj | + (F(t) · n)j,k |Γj,k | = sj (t) |Ωj |, φj (t) ≈ φ(xj , t), (2.4)
k∈N (j)
∂
where φ̇j (t) ≈ ∂t φ(xj , t), |Ωj | is the area of the control volume Ωj and (F(t) ·
n)j,k is the normal component of the numerical flux F(t) on Γj,k , directed from
xj to xk . From now on, we suppress the time dependence of the semi-discrete
equation (2.4) and express it simply as
X
φ̇j |Ωj | + (F · n)j,k |Γj,k | = sj |Ωj |. (2.5)
k∈N (j)
From the above, we get that the spatial discretization of the conservation law
(2.1) requires the approximation of the normal component, F · n, of the numer-
ical flux at the interfaces of the control volume.
For the one-dimensional problem with the flux function given by f = aφ −
νφx , equation (2.5) reduces to
h φ̇i + Fi+1/2 − Fi−1/2 = h si , (2.6)
where Fi+1/2 ≈ f (xi+1/2 ) and h = |Ωi |, see Figure 2.2. Using the second-order
central difference approximation of the derivative φx , the flux approximation
at the interface is given by
ν
Fi+1/2 = a φi+1/2 − (φi+1 − φi ),
h
11
φi
φi−2 φi+1 φi+2

φi−1
Ωi−1 Ωi Ωi+1
Fi−1/2 Fi+1/2
xi−2 xi−1 xi xi+1 xi+2
h
xi−1/2 xi+1/2
Figure 2.2: Cell-centered finite-volume discretization of the 1-D convection-

diffusion equation
where φi+1/2 ≈ φ(xi+1/2 ). Finally, for the flux approximation the unknown φ
needs to be interpolated at the cell-face xi+1/2 of the control volume.
In order to ensure that the resulting discretization has a three-point coup-
ling, φi+1/2 is interpolated using φi and φi+1 , such that
φi+1/2 = c− +
i φi + ci+1 φi+1 ,
with c− + − + − +
i + ci+1 = 1 and 0 ≤ ci ci+1 ≤ 1. The coefficients ci and ci+1 depend on
the approximation scheme used for computing φi+1/2 , e.g., the central scheme
dictates
1
φci+1/2 = (φi + φi+1 ),
2
and thus has c− +
i = 1/2 and ci+1 = 1/2. In Chapter 3, we present the upwind
scheme and local BVP methods for computing φi+1/2 . The numerical flux
Fi+1/2 can now be expressed as
ν
Fi+1/2 = a(c− +
i φi + ci+1 φi+1 ) − (φi+1 − φi )
h
ν
(1 + P c− +

= i )φi + (P ci+1 − 1)φi+1 ,
h
where
ah
P :=
ν
is the mesh Péclet number, a dimensionless parameter giving the ratio of con-
vective and viscous forces. The semi-discrete equation (2.6) can now be formu-
lated as
−ν
− P c− − +

h φ̇i = i−1 + 1 φi−1 + 2 + P (ci − ci ) φi +
h (2.7)
P c+

i+1 − 1 φi+1 + h si .
12
Discrete maximum principle. For robust numerical methods, it is often

considered to be essential that the spatial discretization does not introduce any
spurious oscillations. In [33] it is stated though that spurious oscillations are
not necessarily harmful. They may be useful in signaling that boundary layers
are not resolved and that grid refinement is required. The maximum principle
may help in distinguishing spurious oscillations (spurious local extrema) from
correct oscillations (correct local extrema).
For the homogeneous convection-diffusion problem the maximum principle
states the following.
Theorem 2.1 (Maximum principle). Let φ ∈ C 2 (Ω) ∩ C 0 (∂Ω) and let L be
the differential operator
L[φ] ≡ aφx (x) − νφxx (x) ≤ 0, ∀x ∈ Ω,
then
max φ(x) = max φ(x).
x∈Ω x∈∂Ω
The proof to the above theorem can be found in [15]. Analogous to the max-
imum principle, one can find the minimum principle for L[φ] ≥ 0. Thus, we
get that the homogeneous convection-diffusion equation
aφx − νφxx = 0
can attain its extrema only at the boundary of the domain. Equivalently, the
solution to the above equation does not have a local minimum or maximum in
the interior of the domain.
Next, we present the discrete maximum principle for the discrete operator
Lh defined on the grid xi = x0 + (i − 1)h, i = 1, 2, · · · , N with h = 1/(N − 1),
given by X
Lh [φi ] = ci,k φi+k ,
k∈K
where K is some index set for grid points in the interior of the domain, φi , i =
2, · · · , N − 1.
Definition 2.2. The operator Lh is said to be of positive type, if
X
ci,k = 0, i = 2, · · · , N − 1
k∈K
and
ci,k ≤ 0, k 6= 0, i = 2, 3, · · · , N − 1.
We now present the discrete maximum principle for operators of positive type.
Theorem 2.3 (Discrete maximum principle). If Lh is of positive type and
Lh [φi ] ≤ 0, for i = 2, · · · , N − 1,
then local maxima of φ can occur only if i = 1 or i = N , or if φi = constant.
13
Proof. See chapter 4, [100].

Similarly, if Lh [φi ] ≥ 0 for i = 2, · · · , N − 1 and Lh is of positive type then
it can be shown that the local minima can only occur for i = 1 and i = N .
Thus, if Lh [φi ] = 0 the local extrema can only occur for i = 1 or i = N , i.e., on
the boundary of the domain. Because positive operators do not allow for any
local extremum in the interior of the domain, positive operators do not suffer
from numerical oscillations.
Once we have chosen the coefficients c− +
i and ci+1 , the semi-discrete equation
(2.7) can be expressed as the ODE-system
dΦh
= −AΦh + Sh + b(t), (2.8)
dt
where    
· ·
 ·   · 
   
φi−1  si−1 
   
 φi  ,
Φh =   si  ,
Sh = 
 
φi+1  si+1 
   
 ·   · 
· ·
with A being the coefficient matrix corresponding to the discretization and b
containing the boundary data. The spatial discretization reduces the equation
to a system of ODEs which is numerically integrated using methods described
in the following.
2.1.1 Time integration

The temporal discretization of the system of ODEs (2.8) is done at discrete time
levels tn = n∆t (n = 0, 1, 2, · · · ), with time step ∆t. For the time integration of
the above system we use Runge-Kutta methods, which form a family of one-step
methods. Runge-Kutta methods have been originally developed by Runge [72]
and Kutta [50] with significant contributions by Heun [42]. The integration of
a system
dU
= G(t, U), t > 0, (2.9)
dt
using an s-stage Runge-Kutta method is given by
s
X
Un+1 = Un + ∆t bi k i , (2.10)
i=1
where Un ≈ U(tn ) and ki is defined as

s
X
ki = G tn + ∆t ci , Un + ∆t aij kj , (2.11)
j=1
14
with the coefficients aij , bi and ci given by the Butcher tableau [12]
c1 a11 a12 ··· a1s

c2 a21 a22 ··· a2s
c A .. .. .. .. .. .
= . . . . . (2.12)
b
cs as1 as2 ··· ass
b1 b2 ··· bs
The method is explicit if the matrix A is lower triangular and implicit otherwise.
For the method to be consistent the coefficients in the Butcher tableau must
at least satisfy
s
X
bi = 1, (2.13a)
i=1
s
X
aij = ci for i = 1, 2, · · · , s. (2.13b)
j=1
As the order of the method increases, the number of algebraic conditions on

the coefficients of the method increases. For higher-order methods the pattern
behind the conditions can be more compactly explained using rooted trees, as
shown in [14]. It can be shown that for s-stage methods (s > 4), it is not pos-
sible to have an s-order method with just s stages, with the fifth-order explicit
Runge-Kutta method requiring at least six stages. Explicit Runge-Kutta meth-
ods have a smaller stability region than implicit Runge-Kutta methods. The
implicit Runge-Kutta methods require fewer stages than the explicit methods
for the same order and are much more stable. However, the implicit nature
requires iterative methods to solve the resulting systems of equations if these
are nonlinear, which can significantly increase the computational cost of the
method.
For the convection-diffusion equation, we use the classical fourth-order,
four-stage explicit Runge-Kutta method, given by the Butcher tableau
0
1 1
2 2
1 1
2 2
1 1
1 1 1 1
6 3 3 6
mainly because the higher-order of accuracy allows us to ignore the discretiza-

tion error arising from time integration and, hence, to examine the convergence
of the spatial discretization.
15
For larger time steps implicit-explicit Runge-Kutta methods [2, 25] or impli-
cit Runge-Kutta methods [13, 14] can be used. Alternatively, to have a second-
order accurate time integration scheme that allows for larger time steps, we
can apply the semi-implicit θ-method with θ = 1/2, given by
Un+1 − Un
= θG(tn+1 , Un+1 ) + (1 − θ)G(tn , Un ). (2.14)
∆t
For the ODE-system (2.8) the θ-method translates to
Φn+1 − Φnh
h
= θ −AΦn+1 +Sn+1 +bn+1 +(1−θ) −AΦnh +Snh +bn , (2.15)

h h
∆t
resulting in the discrete system
(I + ∆tθA)Φn+1 = I − ∆t(1 − θ)A Φnh + ∆t θSn+1 + (1 − θ)Snh +

h h
(2.16)
∆t θbn+1 + (1 − θ)bn .

For θ = 1/2, the trapezoidal rule, the time integration is A-stable and second-
order accurate.
Next, we consider the spatial and temporal discretization of the incompress-
ible Navier-Stokes equations.
2.2 Incompressible Navier-Stokes equations

For the spatial discretization of the 2-D incompressible Navier-Stokes equations,
we consider the primitive variable formulation, given by
∇ · u = 0, (2.17a)
∂u 1
+ ∇ · (uu) = −∇p + ∇2 u, = , (2.17b)
∂t Re
where u = u(x, y, t)ex + v(x, y, t)ey , p(x, y, t) and Re represent the velocity
field, the pressure and the Reynolds number of the flow, respectively. The
above system of equations is discretized on a Cartesian staggered grid, with
the pressure and the velocity components defined at different locations on the
grid; see Figure 2.3. The staggering of the variables ensures coupling between
the divergence and the gradient operator. The semi-discrete incompressible
Navier-Stokes equations are given by
Mu = r1 (t), (2.18a)
Ωu̇(t) = −N(u) + Du(t) − Gp(t) + r2 (t), (2.18b)
where u(t) is the discrete velocity vector, and where M, N, D and G represent
the discrete divergence, the non-linear convection, the viscous term and the
gradient operator, respectively. The vector r1 gives the boundary conditions
16
vi,j+3/2
Ωvi,j+3/2
×
pi,j+1 ui+1/2,j+1
vi,j+1/2
Ωui+1/2,j
∆y ×
pi,j ui+1/2,j
∆x
Figure 2.3: Uniform, 2-D staggered grid for the incompressible Navier-Stokes
equations.
17
for the continuity equation (2.17a) and r2 contains the boundary conditions
for the momentum equations (2.17b). The diagonal matrix Ω = diag(Ωu , Ωv )
contains the areas of the u- and v-control volumes (shown in Figure 2.3). Using
the second-order FVM on uniform grids with |Ωu | = |Ωv | = ∆x∆y, the discrete
operators are then given by
(M u)i,j = ūi+1/2,j − ūi−1/2,j + ṽi,j+1/2 − ṽi,j−1/2 , (2.19)
where ūi+1/2,j and ṽi,j+1/2 represent the face-averaged velocity components

computed using the mid-point rule [17], i.e.,
ūi+1/2,j = ∆y ui+1/2,j , ṽi,j+1/2 = ∆x vi,j+1/2 .
The discretization of the nonlinear convective operator N(u) for the u-component
N u is given by
N u (u)i+1/2,j =ūi+1,j ui+1,j − ūi,j ui,j +
ṽi+1/2,j+1/2 ui+1/2,j+1/2 − ṽi+1/2,j−1/2 ui+1/2,j−1/2
(2.20a)
=∆y u2i+1,j − u2i,j + ∆x (uv)i+1/2,j+1/2 −

(uv)i+1/2,j−1/2 .
Details regarding the computation of the cell-face velocity components ui,j , ui+1,j
and (uv)i+1/2,j+1/2 can be found in Chapter 4. Similarly, the v-component of
the convective operator N v is given by
N v (u) i,j+1/2 =ṽi,j+1 vi,j+1 − ṽi,j vi,j + ūi+1/2,j+1/2 vi+1/2,j+1/2 −

ūi−1/2,j+1/2 vi−1/2,j+1/2
2 2
(2.20b)
=∆x vi,j+1 − vi,j + ∆y (uv)i+1/2,j+1/2 −

(uv)i−1/2,j+1/2 .
The discrete diffusion term is given by
(Du (u))i+1/2,j = δx u i+1,j − δx u i,j + δg

y u i+1/2,j+1/2 −
(2.21a)
δg
y u i+1/2,j−1/2 ,
(Dv (u))i,j+1/2 = δx v

i+1/2,j+1/2
− δx v i−1/2,j+1/2 +
(2.21b)
y v i,j+1 − δy v i,j ,
δg g
where
ūi+3/2,j − ūi+1/2,j ũi+1/2,j+1 − ũi+1/2,j
δx u i+1,j = , δg
y u i+1/2,j+1/2 = .
∆x ∆y
Finally, the discrete gradient operator G = (Gu , Gv ) is given by
Gu (p)i+1/2,j = p̄i+1,j − p̄i,j , Gv (p)i,j+1/2 = p̃i,j+1 − p̃i,j . (2.22)
18
The discretization scheme can be extended to achieve higher-order accur-

acy, as shown in [97] for a fourth-order method and to a sixth-order method
as shown in [61]. For both second- and fourth-order methods, the treatment of
the boundary conditions and their conservative properties can be found in [78].
In the incompressible Navier-Stokes equations, pressure, which is a Lagrange
multiplier, ensures that the velocity-field is divergence free and is only determ-
ined up to an additive constant (unless pressure is prescribed at a boundary).
The evolution of the pressure is governed by the pressure-Poisson equation,
which is derived by taking the time-derivative of the continuity equation and
then using the momentum equation, i.e.,
Lp = M F(u, t) − ṙ1 (t), (2.23)
where L = M G is the discrete Laplacian operator and
F(u, t) = −N(u) + Du(t) + r2 (t). (2.24)
In the above expression for F(u, t), Ω−1 has been absorbed by the discrete
operators N, D, G and r2 . The boundary conditions for the pressure-Poisson
equation can be derived using the boundary conditions for the momentum
equations, as described in [35, 79, 94]. Next, the semi-discrete system (2.18) is
integrated using implicit Runge-Kutta methods.
2.2.1 Time integration

For the temporal discretization of the semi-discrete system (2.18) we use the
Runge-Kutta methods as described in [74, 75, 76]. Unlike the semi-discrete
convection-diffusion equation (2.8), which is an ODE system, the semi-discrete
incompressible Navier-Stokes equations constitute a differential-algebraic equa-
tions (DAE) system.
The semi-discrete incompressible Navier-Stokes equations (2.18) can be ex-
pressed as the DAE system [76]
0 =g(u, t), (2.25a)

u̇ =f (u, p, t), (2.25b)
where
g(u, t) =Mu − r1 (t),

f (u, p, t) =F(u, t) − Gp(t),
with F(u, t) as defined in equation (2.24). In an s-stage Runge-Kutta method

the velocity Ui (≈ u(tn + ci ∆t)) at the i-th intermediate stage is determined
by
Xi
Ui = un + ∆t aij Fj − ci GΦi , MUi = r1 (ti ), (2.26)
j=1
19
where aij and ci are the coefficients of the Butcher tableau. Φi is a pressure
like variable satisfying
i
i
X 1 r1 (tn + ci ∆t) − r1 (tn )
LΦ = aij MFj − . (2.27)
c
j=1 i
ci ∆t
In fact, Φi is a Lagrange multiplier to make Ui divergence-free, with each

intermediate velocity field Ui being independent of Φj , j 6= i. Once the
intermediate stage values Ui and Φi are determined, the velocity un+1 is given
by
s
X
ûn+1 = un + ∆t bj Fj , (2.28a)
j=1
s
X r1 (tn+1 ) − r1 (tn )
LΦn+1 = (bj MFj ) − , (2.28b)
j=1
∆t
un+1 = ûn+1 − ∆t GΦn+1 . (2.28c)
We use the second-order Gauss method (implicit mid-point rule) with the
Butcher tableau given by
1 1
2 2
.
1
Details about the implementation of the method and the accuracy of the velo-
city and pressure together with stability analyses are given in [74].
In the following chapters, we present the approximation schemes for the
fluxes involved in the semi-discrete conservation laws.
20
Chapter 3
Convection-diffusion equation
The finite-volume discretization of conservation laws requires fluxes to be ap-

proximated at the interfaces of the control volumes. In this chapter, we present
the discretization of the convective terms in the convection-diffusion equation
using the standard central and upwind schemes and hybrid schemes and also
using two-point local BVPs. For diffusion term, we use the second-order central
approximation. As mentioned in Chapter 1, we require the spatial discretiz-
ation method to be: 3-point (in one space dimension), uniformly second-
order accurate and free of any non-physical oscillations. Higher-order
upwind schemes provide numerical methods that may be stable and at least
second-order accurate. However, such methods can be computationally much
more expensive than the 3-point methods due to wider bandwidths of the dis-
cretization matrix. An extensive analysis of various schemes for the convection-
diffusion equation can be found in, e.g., [19, 26, 27, 68].
In this chapter, we present spatial discretization methods for convection-
diffusion type equations using local boundary value problems (BVPs). The
underlying principle of the scheme is to have a locally exact approximation
of the solution. Our scheme is inspired by the complete-flux scheme for the
convection-diffusion-reaction equation, where the flux is derived from a local
BVP for the entire equation. The integral representation of the flux can be
expressed as the sum of a homogeneous and an inhomogeneous part, corres-
ponding to the homogeneous and particular solution of the BVP, respectively.
We use the complete-flux approach for the discretization of the convective term
and show that the homogeneous approximation, like the hybrid scheme [82],
depends directly on the mesh Péclet number (P ). The homogeneous approx-
imation exhibits the stability properties of the upwind scheme for higher Péclet
numbers and displays second-order accuracy for |P | 1.
To obtain an approximation that is uniformly second-order accurate, it is
essential to include the source terms in the local BVP. For unsteady flow prob-
21
lems, the time derivative of the unknown can be included as part of the source.
In Section 3.3 we show that the inhomogeneous approximation effectively elim-
inates the numerical dissipation from the homogeneous approximation, thereby
resulting in a second-order accurate scheme.
The chapter is organized as follows: in Section 3.1 we present the difference
operators for the spatial discretization of the convection-diffusion equation and
investigate their stability and accuracy. Section 3.2 shows the significance of
the inclusion of the source term in the local BVP, with an analysis of dissipative
properties presented in Section 3.3.
3.1 Convection-diffusion equation

Consider the steady one-dimensional convection-diffusion equation with con-
stant velocity field a(x) = a and constant diffusion constant ν(> 0) for the
unknown φ = φ(x), i.e.,
L[φ] := aφx − νφxx = s, x ∈ (0, L), (3.1)
for some source term s = s(x). For time-dependent problems, the time de-
rivative of the unknown can be included as a source term, i.e, s = −φt . To
numerically solve the above equation, we begin by discretizing the equation
on a uniform grid (xi = ih, i = 0, 1, · · · , N ) with mesh size h = L/N . The
discrete difference operator Lh is then given by
a ν
Lh [φi ] := (φi+1/2 − φi−1/2 ) − 2 (φi−1 − 2φi + φi+1 ) = si , (3.2)
h h
where φi and si are the numerical approximations of φ(xi ) and s(xi ), respect-
ively. In this section, we restrict ourselves to the homogeneous convection-
diffusion equation. In the difference operator Lh , the diffusion term φxx is
approximated by the second-order central difference scheme. The operator Lh
also requires φi+1/2 and φi−1/2 at the auxiliary points xi±1/2 = 21 (xi + xi±1 ),
which should be interpolated using the neighboring values. The central scheme
gives
1
φcen
i+1/2 = (φi + φi+1 ),
2
resulting in the discretization
a ν ν a ν
Lcen
h [φ i ] = − + φ i−1 + 2 φ i + − φi+1
2h h2 h2 2h h2 i (3.3)
ν h
= 2 − 1 + 21 P φi−1 + 2φi − 1 − 12 P φi+1 ,

h
with P = ah/ν the mesh Péclet number.
In Chapter 2 we introduced the notion of positive operators and showed that
they satisfy the discrete maximum principle (Theorem 2.3), and consequently
22
do not allow for non-physical oscillations. Clearly, the operator Lcen

h is positive
if and only if |P | ≤ 2, which imposes a severe restriction on the mesh size for
small values of ν. For smaller ν, convection dominates over diffusion, and in
that case the one-sided discretization, given by the upwind scheme, is suited.
The upwind approximation of φi+1/2 is given by
(
up φi if a ≥ 0,
φi+1/2 =
φi+1 if a < 0,
which results in the discretization

( ν
up h2 − (1 + P )φi−1 + (2 + P )φi − φi+1 if a ≥ 0,
Lh [φi ] = (3.4)
ν

h2 − φi−1 + (2 − P )φi − (1 − P )φi+1 if a < 0.
Clearly, the discretization operator Lup

h is always positive, irrespective of P .
Observe that the upwind approximation of aφx can also be expressed as the
sum of the central approximation and a correction term as
(a 1 ν
2h φi+1 − φi−1 − 2 P h2 φi+1 − 2φi + φi−1 if a ≥ 0,
aφx ≈
a 1 ν

2h φi+1 − φi−1 + 2 P h2 φi+1 − 2φi + φi−1 if a < 0.
From the above, we get that Luph [φi ] gives the same discrete equation as the
central approximation does for the equation

aφx − ν 1 + 12 |P | φxx = 0.
Compared to (3.2), the upwind discretization has a higher diffusion coefficient,

with an additional term νnum = |P |ν/2, which is termed as the numerical
diffusion coefficient. Observe that for |P | 1, we have νnum ν. Numerical
diffusion is an important characteristic of a numerical method. Numerical
diffusion may be introduced to avoid spurious oscillations. However, it may
pollute some subtle features of the flow, e.g., boundary layers which may be
significant, as rehearsed in the following standard example.
Example 3.1. Consider the steady convection-diffusion equation
aφx − νφxx = 0, x ∈ (0, L), a > 0,
subject to the boundary conditions φ(0) = A and φ(L) = B. The analytical

solution of the above equation is given by
1 − ePL (x/L)
φ(x) = A + (B − A) ,
1 − ePL
where PL = aL/ν. This solution has a thin boundary layer when PL is large,
i.e., the solution φ(x) ≈ A all throughout the domain except in a very small
23
φ(x)
φup (x)
1
0.5
0 0.5 1
x
Figure 3.1: Upwind solution compared to the analytical solution for the
convection-diffusion equation with PL = 400.
region near x = L, where the solution attains the boundary condition φ(x) = B.
The thickness of this boundary layer satisfies δ ≈ ν/a = L/PL (see [95]). In
Figure 3.1 we compare the upwind solution with the analytical solution for
PL = 400 and boundary conditions A = 0 and B = 1. Clearly, the upwind
solution has a much thicker boundary layer compared to the analytical solution.
The upwind approximation results in a positive operator, therefore not al-

lowing for any non-physical oscillations, but it introduces numerical diffusion,
which directly depends on the grid Péclet number, leading to loss of details of
the boundary layer.
3.1.1 Homogeneous local BVP approximation

Alternatively, the unknown φi+1/2 can be computed by solving the convection-
diffusion equation (3.1) with s = 0 on (xi , xi+1 ), giving us an approximation
that is locally exact. The scheme was first presented by Allen and Southwell
[18] and has been studied extensively by many others since, see e.g. [33, 66, 82].
The unknown φi+1/2 is thus computed from the local two-point BVP
aφx − νφxx = 0, x ∈ (xi , xi+1 ), (3.5a)

φ(xi ) = φi , φ(xi+1 ) = φi+1 . (3.5b)
24
w+
0.5
w−
-20 -10 0 10 20
P
Figure 3.2: The weights w+ and w− as a function of P .
Using the integrating factor formulation φx − λφ = eλx e−λx φ x , λ = a/ν and

integrating equation (3.5a) once, gives:
− ν(φx − λφ) = −νeλx (e−λx φ)x = c, (3.6)
with c being the integration constant. Next, integrating equation (3.6) from xi
to x ∈ (xi , xi+1 ) and applying the left boundary condition φ(xi ) = φi gives
c x − xi
φ(x) − eP σ(x) φi = 1 − eP σ(x) ,

σ(x) := , a 6= 0,
a h
with σ(x) ∈ [0, 1] being the scaled coordinate on [xi , xi+1 ]. Using the right
boundary condition gives a weighted average of φ on [xi , xi+1 ], i.e.,
e−P (1−σ(x)) − 1 eP σ(x) − 1

φ(x) = φ i + φi+1 . (3.7)
e−P − 1 eP − 1
For a = 0, and consequently P = 0, the above approximation reduces to
φ(x) = (1 − σ(x)) φi + σ(x) φi+1 .
The homogeneous approximation φhi+1/2 = φ(1/2) is given by the weighted

sum
φhi+1/2 = w− φi + w+ φi+1 , w± = W (±P/2), (3.8)
where the weight function W (z) is defined as
1
W (z) := ,
ez + 1
such that W (z) + W (−z) = 1 and 0 < W (z) < 1. Figure 3.2 shows the
dependence of the weights w+ and w− on the grid Péclet number. Note that
for P = 0, w+ = w− = 0.5, resulting in the central approximation, whereas
25
10
1 + P w−
1 − P w+
-10 -5 0 5 10
P
Figure 3.3: Coefficients 1 + P w+ and 1 − P w+ involved in the discrete operator
Lhh
for |P | > 10, φhi+1/2 ≈ φup

i+1/2 . Alternatively, the homogeneous approximation
h
φi+1/2 can be expressed as a weighted sum of the upwind and the central
approximation, given by
up
φhi+1/2 = 2W (|P |/2) φcen
i+1/2 + (1 − 2W (|P |/2)) φi+1/2 . (3.9)
The homogeneous local BVP approximation results in the discretization

ν
Lhh [φi ] = −(1+P w− )φi−1 + 2+P (w− −w+ ) φi −(1−P w+ )φi+1 . (3.10)

h2
For Lhh to be a positive operator we require 1+P w− and 1−P w+ to be positive.

Figure 3.3 shows a plot of the coefficients 1 + P w− and 1 − P w− as a function
of the grid Péclet number.
Lemma 3.1. The difference operator Lhh is positive for all P ∈ R.
Proof. Consider the coefficient
eP/2 + 1 − P
1 − P w+ = .
eP/2 + 1
Clearly eP/2 + 1 is always positive. For 1 − P w+ to be positive, it is thus

necessary that eP/2 − P + 1 > 0. To show this, we define the function
C1 (z) := ez/2 − z + 1,
with (C1 )z = ez/2 /2 − 1 and (C1 )zz = ez/2 /4 > 0. Thus, the function C1 (z)
has a minimum at z = 2 log(2), given by 3 − log 2 > 0.
Next, for the coefficient 1 + P w− to be positive we need to prove that
1 + P w− > 0, or 1 + eP/2 (1 + P ) > 0.
26
The above condition holds trivially for P ≥ −1. For P < −1, we define the
function
C2 (z) := ez/2 (1 + z).
Differentiating C2 (z) with respect to z, we get that C2 (z) has a global minimum
at z = −3, with C2 (−3) = −2e−3/2 ≈ −0.45. Therefore, 1 + C2 (z) > 0, for all
z ∈ R, resulting in 1 + P w− being positive.
The discretization schemes described above can be expressed as a linear
system
AΦh = b,
where A is the discretization matrix corresponding to the discrete operator Lh
and where b contains the information about the boundary conditions. Thus
we have the discretization matrices Acen , Aup and Ah corresponding to the
up
difference operators Lcen h
h , Lh and Lh , respectively.
Of the above 3-point discretization schemes for equation (3.1), the homo-
geneous local BVP approximation and the upwind scheme always result in a
positive operator, whereas the central scheme is positive only if |P | ≤ 2.
Local discretization error. The local discretization error τi for the discrete
difference operator Lh [φi ] is defined as
Lh [φ(xi )] = si + τi .
Using Taylor series expansions, the derivatives involved in Lh [φ(xi )] can be

expressed as
dφ 1 h d2 φ
(xi ) = φ(xi ) − φ(xi−1 ) + (xi ) + O(h2 ), (3.11a)
dx h 2 dx2
2
h d φ 3
dφ 1
(xi ) = φ(xi+1 ) − φ(xi−1 ) − (xi ) + O(h4 ), (3.11b)
dx 2h 6 dx3
d2 φ 1 h2 d4 φ
(xi ) = φ(xi+1 ) − 2φ(xi ) + φ(xi−1 ) − (xi ) + O(h4 ). (3.11c)
dx2 h 2 12 dx4
Using (3.4), (3.11a) and (3.11c), it can be shown that the upwind scheme is
first-order accurate with
ah d2 φ h2 d4 φ
τiup = − 2
(xi ) − ν (xi ) + O(h3 ), a ≥ 0.
2 dx 12 dx4
Similarly, using (3.11b), (3.11c) and (3.3) it can be shown that the central
scheme is second-order accurate, with the local discretization error given by
h2 d3 φ ν d4 φ
τicen = a 3 (xi ) − (xi ) + O(h4 ).
6 dx 2 dx4
The homogeneous local BVP approximation φhi+1/2 is given by the linear com-
bination (3.9) of the upwind and the central schemes. Consequently, the local
27
discretization error for the homogeneous local BVP scheme can be expressed as
the same linear combination of the corresponding discretization errors, given
by
1 d2 φ h2 d3 φ ν d4 φ
τih = − ah(1−2w+ ) 2 (xi )+ 2aw+ 3 (xi )− (x 3
i +O(h ),
) a ≥ 0.
2 dx 6 dx 2 dx4
On coarse grids, i.e., for higher Péclet numbers (P > 10), the homogeneous
local BVP scheme reduces to the upwind scheme and exhibits first-order con-
vergence. For finer grids, it reduces to the central scheme and exhibits second-
order convergence.
We have established that the homogeneous local BVP scheme results in a
positive discretization operator and is second-order accurate over finer grids.
The accuracy of the local BVP method can be further improved by the inclusion
of the source term in the model BVP (3.5), as demonstrated in [55] for the
unsteady constant-coefficient convection-diffusion-reaction equation, where the
authors have shown that the inclusion of the source terms makes the complete
flux scheme uniformly second-order convergent.
3.2 Inhomogeneous local BVP approximation

To obtain a more accurate approximation of φi+1/2 , we can take into account
the source term in the local BVP for the approximation. To demonstrate the
significance of the source term, we consider the inhomogeneous local BVP
aφx − νφxx = s, x ∈ (xi , xi+1 ), (3.12a)
φ(xi ) = φi , φ(xi+1 ) = φi+1 . (3.12b)
The linearity of the problem allows us to express the solution of the inhomo-
geneous problem as the sum of the homogeneous and the inhomogeneous part,
as follows
φi+1/2 = φhi+1/2 + φinh
i+1/2 ,
with φhi+1/2 corresponding to the solution of the homogeneous problem given

by (3.8) and φinhi+1/2 being the particular solution corresponding to the source
term s. The solution of the above inhomogeneous BVP can be evaluated under
the assumption that the source term s is either constant or piecewise constant
on (xi , xi+1 ). In this chapter, we only consider a constant source, the details
for the derivation of the solution to BVP (3.12) for piecewise constant source
term can be found in Chapter 4.
3.2.1 Source term for unsteady flow problems

Consider the unsteady convection-diffusion equation
φt + aφx = νφxx , (x, t) ∈ (0, L) × [0, T ], (3.13)
28
with initial condition φ(x, 0) = φ0 (x) and the boundary conditions φ(0, t) =
A(t), φ(L, t) = B(t) and φ0 (0) = A(0), φ0 (L) = B(0). Analogous to equation
(3.2), the above equation can be discretized as
a ν
φ̇i + (φi+1/2 − φi−1/2 ) − 2 (φi−1 − 2φi + φi+1 ) = 0. (3.14)
h h
For the approximation of φi+1/2 , we solve the inhomogeneous quasi-steady BVP
1
(φx − λφ)x = φ̇i+1/2 , xi < x < xi+1 , (3.15a)
ν
φ(xi ) = φi , φ(xi+1 ) = φi+1 , (3.15b)
with source term given by the upwind value of the discrete time derivative of
the unknown, i.e.,
(
φ̇i if a ≥ 0,
φ̇i+1/2 =
φ̇i+1 if a < 0.
Substituting the integrating factor formulation φx − λφ = eλx (e−λx φ)x in the

differential equation and integrating gives
1 −λx
(e−λx φ)x =

e φ̇i+1/2 (x − xi+1/2 ) + k1 ,
ν
with k1 the integration constant. Integrating the above equation from xi to
x ∈ (xi , xi+1 ) gives
Z x
1
φ(x) =eλ(x−xi ) φi + φ̇i+1/2 eλ(x−ξ) (ξ − xi+1/2 )dξ+
ν xi
Z x
1
k1 eλ(x−ξ) dξ.
ν xi
Using the normalized coordinate σ(x), the solution of the local BVP (3.15) can
be expressed as
σ
h2
Z
k1 P σ
φ(σ) = eP σ φi + eP (σ−η) η − 1

φ̇i+1/2 2 dη + (e − 1), a 6= 0. (3.16)
ν 0 a
Note that for a = 0 (equivalently P = 0), the above expression reduces to
h2
φ(σ) = (1 − σ)φi + σφi+1 − σ(1 − σ)φ̇i+1/2 .
2ν
The constant k1 in equation (3.16) is determined by imposing the boundary
condition φ(xi+1 ) = φi+1 . To compute φi+1/2 , we substitute σ = 1/2, which
29
0.5
tanhc(z/4)
0
-20 -10 0 10 20
z
Figure 3.4: Function tanhc(z/4).
gives
φi+1/2 =φhi+1/2 + φinh
i+1/2 , (3.17a)
−
φhi+1/2 +
=w φi + w φi+1 , (3.17b)
1 2 tanh(z)
φinh
i+1/2 =− h tanhc 41 P φ̇i+1/2 , tanhc(z) = . (3.17c)
8ν z
Figure 3.4 shows the plot of tanhc(z). It can be observed that tanhc(P/4)
attains its maximum for P = 0 and then asymptotically goes to 0 as |P | → ∞.
For the sake of simplicity, we restrict ourselves to the case a ≥ 0, thus
φ̇i+1/2 = φ̇i . Using the above approximation in the semi-discrete convection-
diffusion equation (3.14) results in the ODE-system
(Ω + B) Φ̇h = −Ah Φh + b, (3.18)
where Ω = hI is the mass matrix and where B is given by
 
−1
 1 −1 

. .
 1 2
B = aβ  .. .. , β = h tanhc 14 P .
 
  8ν
 1 −1 
1 −1
Thus, the inclusion of the time derivative as a source term in the local BVP
results in a modified mass matrix Ω + B.
Example 3.2. Consider the unsteady convection-diffusion problem

φt + aφx − νφxx = 0, (x, t) ∈ (−1, 1) × [0, 1],
30
10 0
kekcentral
kekupwind
kekhom
10 -1 kekinh
||e||∞
10 -2
2
10 -3
10 -4
10 -3 10 -2 10 -1 10 0
h
Figure 3.5: Convergence of the spatial errors for a = 1 and ν = 10−3 for a
family of grids with h = 2−i × 10−1 , i = −1, 0, 1, 2, 3, 4 and ∆t = 10−4 .
with initial conditions φ(x, 0) = sin(πx) and boundary conditions

φ(−1, t) = φ(1, t) = sin(π at) exp(−π 2 νt).
For the above problem, the analytical solution is given by
φ(x, t) = sin(π(x − at)) exp(−π 2 νt),
which describes a decaying traveling wave. We set a = 1.0 and ν = 10−3 and
compare the numerical results obtained using the upwind, central, homogeneous
local BVP and the inhomogeneous local BVP scheme. We use the standard,
explicit fourth-order, four-stage Runge-Kutta method with ∆t = 10−4 , ensur-
ing that the temporal errors are negligible compared to the spatial errors. Fig-
ure 3.5 shows the convergence of the absolute errors (kek∞ = kφexact −φnum k∞ ,
with φexact being the analytical solution restricted to the grid) for different spa-
tial discretization schemes. Due to the relatively high Péclet numbers, ranging
from 12.5 to 400, the homogeneous local BVP approximation reduces to the
upwind scheme and exhibits first-order convergence. Further, the high grid
Péclet numbers cause the central scheme to suffer from non-physical oscilla-
tions, thereby inhibiting uniformly second-order convergence. Lastly, the in-
homogeneous approximation (depicted by kekinh ) results in a uniformly second-
order accurate scheme.
In the next section, we investigate the numerical dissipation of the ho-
mogeneous and inhomogeneous approximations, to explain the second-order
convergence of the inhomogeneous local BVP scheme.
31
3.3 Numerical dissipation and dispersion

To understand the effect of the inclusion of the source term, i.e., the time
derivative of the unknown, we analyze numerical dissipation and dispersion
of the homogeneous and the inhomogeneous approximation schemes applied
to the unsteady convection-diffusion equation. For the following analysis we
derive the modified equation for both schemes, as described in [88] for the
complete-flux scheme applied to the convection-diffusion-reaction equation. To
do so, we replace φni by φ(xi , tn ) in the discrete equation and then expand all
nodal values of φ(x, t) in a Taylor series about (xi , tn ). Next, we eliminate the
higher-order derivatives (2nd and above) using the modified equation itself, as
described in [98].
We begin by deriving the modified equation for the homogeneous local BVP
method. Consider the difference operators
1 1
δx φi := (φi+1 − φi−1 ), δx− φi := (φi − φi−1 ),
2h h
1
δxx φi := 2 (φi+1 − 2φi + φi−1 ),
h
defined for a grid function φi . Using the homogeneous approximation (3.8) for
the interface values φi+1/2 and φi−1/2 results in the difference equation
a + ν
w φi+1 + (w− − w+ )φi − w− φi−1 − 2 φi+1 − 2φi + φi−1 = 0. (3.19)

φ̇i +
h h
Upon rearranging the convective term, equation (3.19) can be expressed as the
sum of the central approximation and a correction term as follows,
a ν
φi+1 − φi−1 − 2 1 + P 12 − w+

φ̇i + φi+1 − 2φi + φi−1 . (3.20)
2h h
Thus, the homogeneous scheme is equivalent to the semi-discretization

h h
φ̇i + aδx φi − νart δxx φi = 0, νart = ν 1 + P 12 − w+ , (3.21)
h
where νart is the artificial diffusion coefficient, defined as the sum of the diffusion
coefficient ν and the numerical diffusion
h
νnum = 2νD(P/4), D(z) := z tanh(z).
h
Figure 3.6 shows the plot of νnum /ν, as a function of P , from which it can be ob-
served that the homogeneous approximation introduces less numerical diffusion
than the upwind scheme. Clearly, for P = 0, the homogeneous approximation
does not introduce any numerical diffusion, whereas, for higher Péclet numbers
h
(|P | 1), νnum ≈ |P |ν/2, the numerical diffusion from the upwind scheme.
32
2
Homogeneous
1.5 Upwind
0.5
0
-4 -2 0 2 4
P
Figure 3.6: Plot of the numerical diffusion νnum /ν for the homogeneous local
BVP approximation (νnum /ν = 2D(P/4)) and the upwind scheme (νnum /ν =
|P |/2).
The modified equation corresponding to the semi-discretization (3.21) reads

2
∂ψ ∂ψ h ∂ ψ
+a = νart . (3.22)
∂t ∂x ∂x2
Thus, the homogeneous approximation suffers from numerical diffusion.
However, the scheme does suppress any non-physical spurious oscillations, as
the discrete operator Lhh is always positive.
Next, we consider the inhomogeneous approximation scheme with s =
−φ̇i+1/2 , which is assumed to be constant on (xi , xi+1 ) and is given by the
upwind value of the discrete time derivative. For a ≥ 0, φi+1/2 is computed
from the inhomogeneous local BVP (3.15) and φi+1/2 is expressed as the sum
of a homogeneous and an inhomogeneous part given by (3.17). Substituting
the expressions for φi+1/2 and φi−1/2 in the semi-discrete convection-diffusion
equation (3.14) gives
h
I − tanh(P/4)δx− φ̇i + aδx φi − νart h
δxx φi = 0, (3.23)
2
where I is the identity operator. The inverse of I − (h/2)tanh(P/4)δx− is given
by the relation
h −1 1 h −1 1 h
I − tanh(P/4)δx− = I − νnum δx− = I + νnum δx− + O(h2 ). (3.24)
2 a a
Applying the inverse to equation (3.23) results in
h h
δx− δx φi −

φ̇i + aδx φi − ν + νnum δxx φi + νnum
1 h h
−
ν ν + νnum δx δxx φi = O(h2 ).
a num
33
(a) (b)
20 2
D(z)
E(z)
10 1
0 0
-20 -10 0 10 20 -20 -10 0 10 20
z z
Figure 3.7: Functions (a) D(z) and (b) E(z).
Observe that the extra diffusive term introduced by applying the inverse to the
convective term, i.e.,
h
νnum δx− δx φi = νnum
h
δxx φi + O(h2 ),
exactly cancels the numerical diffusion arising from the homogeneous part of
the approximation. Further, applying the inverse to the diffusive term gives
rise to the dispersive term
h
(1/a)νnum h
(ν + νnum )δx− δxx φi ,
with δx− δxx φi being the first-order approximation of (∂ 3 φ/∂x3 )(xi ). The semi-
discrete equation (3.23) can be expressed as
φ̇i + aδx φi = νδxx φi + δx− δxx φi + O(h2 ), = ah2 E(P/4), (3.25)
where is the numerical dispersion coefficient and
1 2
E(z) := 2tanhc(z) + tanh(z) .
4
Equation (3.25) is thus a second-order accurate approximation of the equation
∂ψ ∂ψ ∂2ψ ∂3ψ
+a =ν 2 + 3, (3.26)
∂t ∂x ∂x ∂x
with ∂ 3 ψ/∂x3 a dispersion term, causing waves of different wavelengths to
travel at different speeds [103]. Figure 3.7 shows the functions D(z) and E(z),
E(z) is a monotonically decreasing function with
lim E(P ) = 1/4.
|P |→∞
34
To demonstrate the effect of the dispersion term in equation (3.26), we consider

a planar wave in the inviscid case. Note that in the absence of diffusion, i.e.,
ν = 0, or equivalently, |P | → ∞. Substituting the planar wave
ψ(x, t) = Aei(κx−ωt) ,
with amplitude A, frequency ω and wave number κ, into equation (3.26) (with
ν = 0) gives us the dispersion relation
ω(κ) = aκ + κ3 6= const.
The phase velocity corresponding to the dispersion relation is then given by
ω(κ) 1
cp (κ) = = a + κ2 = a 1 + (hκ)2 > a,
κ 4
causing the numerical solution to propagate faster.
Thus, including the source term, s = −φ̇ in the local BVP, eliminates the
numerical diffusion arising from the homogeneous part of the approximation,
at the cost of introducing a dispersion error.
3.4 Conclusion
In this chapter, we introduced the discretization of the convective term in the
convection-diffusion equation using local BVPs. Integrating the local BVP
provides us the integral representation, which can be expressed as the sum of a
homogeneous and an inhomogeneous part, corresponding to the homogeneous
and particular solution of the local BVP, respectively. The homogeneous ap-
proximation, is a weighted sum of the central and the upwind scheme, such that
the homogeneous approximation reduces to the upwind scheme for |P | → ∞
and to the central scheme for |P | 1. Consequently, it exhibits second-order
convergence on finer grids and first-order on coarse grids. To have a uniformly
second-order scheme it is essential to include the inhomogeneous part of the
approximation. We have shown that inclusion of the source term s = −φt
cancels the numerical diffusion introduced by the homogeneous part of the
approximation.
In the next chapter, we demonstrate the use of local BVPs for the discret-
ization of the convective terms in the incompressible Navier-Stokes equations,
with the pressure gradient and the gradient of the transverse flux forming the
source term. Like for the convection-diffusion equation, where the inclusion
of the source term makes the discretization second-order accurate, we show
that inhomogeneous local BVP methods applied to the Navier-Stokes equation
result in second-order convergent schemes.
35
Chapter 4
Convective discretization of the

Navier-Stokes equations
In Chapter 2, we discussed the spatial discretization of the Navier-Stokes equa-

tions using staggered grids. In order to evaluate the semi-discrete convective
operator N, velocity components at the interfaces of the control volumes are
required. Commonly used methods for the approximation of these cell-face ve-
locity components are the central and the upwind scheme. The central scheme
approximates the velocity components by averaging the neighboring grid point
velocity components. The upwind scheme gives an approximation of the velo-
city components depending on the direction of the flow. Both methods have
their trade-offs. The central scheme does not add any non-physical dissipa-
tion to the system, but may suffer from spurious oscillations, particularly at
higher Reynolds numbers [36, 105]. The upwind scheme yields good stability
properties [10], but adds significant numerical dissipation to the system. An
analysis of the energy-conservation properties of the two schemes can be found
in [61]. The above two methods for approximation of the cell-face velocities are
the limit-case methods for the momentum equations, i.e., the central scheme
corresponds to the no-flow situation, whereas the upwind scheme corresponds
to inviscid flow situation. Higher-order approximation of the cell-face velocity
components using for example (W)ENO reconstruction provides a more accur-
ate numerical solution that is free of any spurious oscillations. However, these
discretization methods are computationally more expensive than the convec-
tional 3-point (in each spatial direction) discretization methods.
An accurate approximation of the cell-face velocities should take into ac-
count the balance of the convective and viscous forces acting on the fluid. This
issue can be tackled by the exponential/hybrid scheme described in [63, 82],
which gives an approximation of the cell-face velocities as a weighted average
of the central and the upwind scheme, with the weight function depending on
37
Convective discretization of the Navier-Stokes equations
the mesh Péclet number (ratio of the convective and viscous forces).
Apart from the convective and viscous forces, the cell-face velocities are
also affected by the pressure gradient and the effect of rotational behavior
(quantitatively given by the gradient of the transverse flux) of the fluid flow.
In this chapter we present methods for the computation of cell-face velocities
using reduced momentum equations subject to suitable boundary conditions,
reducing the problem to a local boundary value problem (BVP). We aim to
have approximation methods for cell-face velocities, that are accurate, free of
numerical oscillations, introduce no or very little numerical dissipation and
result in a 3-point coupling (in each spatial direction).
The chapter is organized as follows: In Section 4.1 the finite-volume method
is briefly described, giving a mathematical formulation for the computation of
cell-face velocity components. In Section 4.2, we derive an integral represent-
ation for the cell-face velocities by solving local BVPs. Sections 4.3 and 4.4
give the details about the closure of the model and iterative computation of
the cell-face velocities. The numerical results for the local BVP methods are
discussed in Section 4.5 and conclusions are drawn in Section 4.6.
4.1 Convective discretization

Consider the two-dimensional incompressible Navier-Stokes equations in prim-
itive variable formulation, i.e.,
∇ · u = 0, (4.1a)
∂u 1 2
+ ∇ · (uu) = −∇p + ∇ u. (4.1b)
∂t Re
As shown in Chapter 2 the above system of equations is discretized as
Mu = r1 (t), (4.2a)
1
Ωu̇(t) = −N(u) + Du(t) − Gp(t) + r2 (t), (4.2b)
Re
the details of the finite-volume discretization can be found in Section 2.2. For
the control volume Ωui+1/2,j (see Figure 4.1), the u-component N u of the dis-
cretized convection operator N = (N u , N v ) is given by
N u (u) i+1/2,j =∆y u2i+1,j − u2i,j +

(4.3)
∆x (uv)i+1/2,j+1/2 − (uv)i+1/2,j−1/2 ,
where (uv)i+1/2,j+1/2 = ui+1/2,j+1/2 vi+1/2,j+1/2 and ui,j is the numerical ap-
proximation of the velocity component u(xi , yj ). Similarly, for the v-control
volume Ωvi,j+1/2 , the discrete convective operator (N v (u))i,j+1/2 is given by

N v (u) i,j+1/2 =∆y (uv)i+1/2,j+1/2 − (uv)i−1/2,j+1/2 +

(4.4)
2 2
∆x vi,j+1 − vi,j .
38
Ωvi,j+1/2
pi,j+1
yj+1 o
Ωui+1/2,j
vi,j+1/2
∆y pi,j
yj o o
ui+1/2,j ui+3/2,j
xi xi+1
∆x
Figure 4.1: Control volumes for the spatial discretization of the momentum
equations.
In order to compute (N u (u))i+1/2,j we need the velocity components ui+1,j ,

ui+1/2,j+1/2 and vi+1/2,j+1/2 , the rest can be computed in a similar manner.
These velocity components are not defined on the grid structure used for the
discretization, and hence need to be approximated using neighboring grid ve-
locities.
We begin with the approximation of the velocity component ui+1,j , the
approximation of the other cell-face velocity components is described in Sec-
tion 4.3. The central approximation of the velocity component ui+1,j is given
by
1
ui+1,j = (ui+1/2,j + ui+3/2,j ).
2
The upwind approximation depends on the direction of the flow and is defined
as

ui+1/2,j ,
 if ui+1/2,j > 0 ∧ ui+3/2,j > 0,
ui+1,j = ui+3/2,j , if ui+1/2,j < 0 ∧ ui+3/2,j < 0,
1

2 (ui+1/2,j + ui+3/2,j ), if ui+1/2,j ui+3/2,j ≤ 0.
Alternatively, we compute the velocity component ui+1,j by solving the u-

momentum equation locally, such that the computed cell-face velocity com-
ponent depends on the mesh Péclet number and source terms like the pressure
gradient and the cross-flux gradient. Consider the steady u-momentum equa-
tion, i.e.,
1
(u2 )x + (uv)y = −px + (uxx + uyy ).
Re
Restricting the solution of the above equation to the domain x ∈ (xi+1/2 , xi+3/2 ),
39
(δx p)i+1/2,j (δx p)i+3/2,j

ui+1,j
yj × × ×
ui+1/2,j ui+3/2,j
u
Ci+1,j
xi+1/2 ∆x xi+3/2
Figure 4.2: Control-volume for the computation of ui+1,j
y = yj , gives us the local BVP
(u2 − ux )x = −px − (uv − uy )y , = 1/Re, (4.5a)

u(xi+1/2 , yj ) = ui+1/2,j , u(xi+3/2 , yj ) = ui+3/2,j . (4.5b)
Figure 4.2 shows the domain and the boundary conditions involved in the
computation of ui+1,j , the details of which are given in the following section.
4.2 Integral representation of cell-face

velocities
In this section we focus on the computation of ui+1,j from the local two-point
BVP (4.5). Equation (4.5a) is a non-linear, inhomogeneous partial differential
equation, which does not allow for exact representation of the solution. To
simplify the problem, the non-linear term (u2 )x is linearized as U ux where U is
an approximation of the velocity component ui+1,j . We assume that the gradi-
ent of the cross-flux (uv − uy )y is constant over the interval (xi+1/2 , xi+3/2 )
u
and replace it with its numerical approximation Ci+1,j ≈ (uv − uy )y |(xi+1 ,yj ) .
Further, we assume that the pressure p has a piecewise linear profile over
(xi+1/2 , xi+3/2 ), y = yj , with a kink at x = xi+1 , as shown in Figure 4.3, mak-
ing the (discrete) pressure gradient δx p piecewise constant. With the above
assumptions, equation (4.5a) reduces to a linear, inhomogeneous, quasi-one-
dimensional ordinary differential equation, given by
u
(U u − ux )x = −δx p − Ci+1,j , xi+1/2 < x < xi+3/2 , y = yj . (4.6)
In the following derivation we suppress the y-dependence of the local BVP,

u(xi ) := u(xi , yj ), and introduce the following variables,
a := U/, P := a ∆x ,
40
pi+1,j
pi+2,j
pi,j
xi+1/2 xi+3/2
Figure 4.3: The piecewise linear pressure profile over the domain of the local
BVP.
where ∆x = xi+3/2 − xi+1/2 and P is the mesh Péclet number. Thus equation
(4.6) can be rewritten as
1 u
(ux − au)x = (δx p + Ci+1,j ). (4.7)

Since the pressure is assumed to be piecewise linear, we get that the pressure
gradient px is given by
1
(
(δx p)i+1/2,j = ∆x (pi+1,j − pi,j ), xi+1/2 ≤ x ≤ xi+1 ,
δx p(x) =
1
(δx p)i+3/2,j = ∆x (pi+2,j − pi+1,j ), xi+1 < x ≤ xi+3/2 .
Integrating equation (4.7) we get

Z x
1 u

ux − au = I1 (x) + Ci+1,j (x − xi+1 ) + k1 , I1 (x) = δx p(ξ)dξ , (4.8)
xi+1
where k1 is the integration constant. Note that we integrated equation (4.7)

from xi+1 to x ∈ (xi+1/2 , xi+3/2 ), as the pressure gradient px has a jump at
x = xi+1 . Integrating equation (4.8) using the integrating factor formulation,
ux − au = eax (e−ax u)x , and invoking the left boundary condition u(xi+1/2 ) =
ui+1/2 , we get
Z x
1
u(x) =ea(x−xi+1/2 ) ui+1/2,j + ea(x−ζ) I1 (ζ) dζ+
xi+1/2
Z x Z x (4.9)
1 u a(x−ζ) 1 a(x−ζ)
C e (ζ − xi+1 ) dζ + k1 e dζ.
i+1,j xi+1/2 xi+1/2
We define the normalized coordinate σ on the interval [xi+1/2 , xi+3/2 ] by
x − xi+1/2
σ := σ(x) = .
∆x
41
Using equation (4.9) and the normalized coordinate σ we get

1 1 u k1
u(σ) = eP σ ui+1/2,j + ∆x I2 (σ; P ) + Ci+1,j ∆x2 I3 (σ; P ) + ∆x σϕ(P σ),

where ϕ(z) = (ez − 1)/z and where the integral functions I2 and I3 are defined
as
Z σ
I2 (σ; P ) := eP (σ−η) I1 (xi+1/2 + η∆x) dη,
0
Z σ
eP (σ−η) η − 12 dη.

I3 (σ; P ) :=
0
Evaluating the integral I3 (σ; P ) gives

1
I3 (σ; P ) = 2 1 − 12 P eP σ − 1 − σP .

(4.10)
P
Invoking the right boundary condition u(xi+3/2 ) = ui+3/2,j we get
1
u(σ) =H1 (1 − σ; −P )ui+1/2,j + H1 (σ; P )ui+3/2,j + ∆x I2 (σ; P )−

1
u 2
H1 (σ; P )I2 (1; P ) + Ci+1,j ∆x I3 (σ; P ) − H1 (σ; P )I3 (1; P ) ,

(4.11)
where
eP σ − 1
H1 (σ; P ) := .
eP − 1
Summarizing, the integral representation u(σ) can be expressed as the sum of
three parts, i.e.,
u(σ) = uh (σ) + up (σ) + uc (σ), (4.12a)

h
u (σ) := H1 (1 − σ; −P )ui+1/2,j + H1 (σ; P )ui+3/2,j , (4.12b)
1
up (σ) := ∆x I2 (σ; P ) − H1 (σ; P ) I2 (1; P ) , (4.12c)

1 u
u (σ) := Ci+1,j ∆x2 I3 (σ; P ) − H1 (σ; P ) I3 (1; P ) ,
c
(4.12d)

with uh (σ) corresponding to the homogeneous solution of the local BVP (4.7)
and up (σ), uc (σ) to the discrete pressure gradient and the gradient of the
cross-flux, respectively. Note that uh , up and uc satisfy the following boundary
conditions:
uh (0) = ui+1/2,j , uh (1) = ui+3/2,j , (4.13a)

p p
u (0) = u (1) = 0, (4.13b)
uc (0) = uc (1) = 0. (4.13c)
42
In order to compute ui+1,j , we set σ = 21 . Using the weight function

W (z) = (ez + 1)−1 , satisfying 0 ≤ W (z) ≤ 1 and W (z) + W (−z) = 1, uh ( 21 )
can be expressed as the weighted average of ui+1/2,j and ui+3/2,j :
uh ( 12 ) = W (−P/2)ui+1/2,j + W (P/2)ui+3/2,j . (4.14)
Computing the integral I2 (σ; P ) for σ = 1/2 and σ = 1, we get

1
I2 12 ; P = ∆x pi+1,j − pi,j 2 1 − 21 P eP/2 − 1 ,

P
I2 (1; P ) = e I2 2 ; P + ∆x pi+2,j − pi+1,j eP/2 − 1 − 12 P .
P/2 1

Introducing the function
eP − 1 − P
H2 (P ) := ,
P 2 (eP + 1)
1
and I2 (1; P ), up ( 12 ) can be written as

and using the expressions for I2 2; P
1
up 1
∆x2 H2 (−P/2)(δx p)i+1/2,j + H2 (P/2)(δx p)i+3/2,j .

2 =− (4.15)
4
1

Finally, uc 2 can be expressed as
1 P
u
uc 1
∆x2 Ci+1,j

2 =− tanhc . (4.16)
8 4
Observe that the function H2 (z) and tanhc(z) satisfy
1
H2 (−z) + H2 (z) = tanhc(z/2), (4.17)
2
implying that inhomogeneous part (4.15) with piecewise constant source term
is equivalent to expression (4.16), if (δx p)i+1/2,j = (δx p)i+3/2,j . Summarizing,
we get that the velocity component ui+1,j is given by
ui+1,j = uh ( 21 ) + up ( 12 ) + uc ( 12 ), (4.18)
with uh ( 12 ), up ( 12 ) and uc ( 12 ) defined in (4.14), (4.15) and (4.16), respectively.

Depending on the terms in the right-hand side of equation (4.7) we get three
possible methods for computing ui+1,j , viz., the homogeneous 1-D local BVP
method (uh ), the inhomogeneous 1-D local BVP method (uh + up ) and the 2-D
local BVP method (uh + up + uc ), specified in the following.
4.2.1 Homogeneous 1-D local BVP method

If we ignore the pressure gradient and the gradient of the cross-flux, then solving
the homogeneous equation
(ux − au)x = 0 (4.19)
43
0.25
0.15
H2 (−z/2)
0.05
H2 (z/2)
-20 -10 0 10 20
z
Figure 4.4: Function H2 .
gives the cell-face velocity component ui+1,j as a weighted average of ui+1/2,j

and ui+3/2,j as in equation (4.14). This is the homogeneous 1-D local BVP
method. The homogeneous part can be expressed as a weighted average of the
upwind and the central scheme:
uh ( 21 ) = (1 − 2W (|P |/2))uup + 2W (|P |/2)ucen . (4.20)
4.2.2 Inhomogeneous 1-D local BVP method

The inclusion of the pressure gradient results in the inhomogeneous differential
equation,
1
(ux − au)x = δx p, (4.21)

in which case the cell-face velocity component is expressed as ui+1,j = uh ( 12 ) +
up ( 21 ). The contribution of the discrete pressure gradients (δx p)i+1/2,j and
(δx p)i+3/2,j to the component up is controlled by the function H2 . Figure
4.4 gives the dependence of H2 (P/2) on the Péclet number. It can be seen
that for P = 0, (δx p)i+1/2,j and (δx p)i+3/2,j have equal contribution. How-
ever, as the Péclet number (P > 0) increases, we get that H2 (−P/2) → 2/P
and H2 (P/2) → 4/P 2 , as P → ∞, making the upwind pressure gradient
(δx p)i+1/2,j more dominant. Similarly for P → −∞, the discrete pressure
gradient (δx p)i+3/2,j is more dominant than (δx p)i+1/2,j .
4.2.3 2-D local BVP method

Finally, including both the pressure gradient and cross-flux gradient and solving
differential equation (4.7), results in the cell-face velocity component given by
44
equation (4.18). The term uc ( 12 ) is proportional to the cross-flux gradient Ci+1,j

u
and the function tanhc(P/4) (see Figure 3.4), which attains its maximum for
P = 0 and then asymptotically goes to 0 as |P | → ∞.
Approximation of the gradient of the cross-flux

In order to compute the term uc (1/2) given by equation (4.16), we need the
u u u u
cross-flux gradient Ci+1,j . Ci+1,j is computed using Ci+1/2,j and Ci+3/2,j ,
which are approximations of the gradient (uv − uy )y at (xi+1/2 , yj ) and
u
(xi+3/2 , yj ), respectively, with Ci+1/2,j defined as
u 1
Ci+1/2,j = (uv)i+1/2,j+1/2 − (uv)i+1/2,j−1/2 −
∆y

2
ui+1/2,j+1 − 2ui+1/2,j + ui+1/2,j−1 .
∆y
u
The gradient Ci+3/2,j can be computed similarly. For the inhomogeneous 1-D
local BVP approximation, we have seen that for |P | 1, the upwind pressure
gradient has a higher contribution and for |P | 1, both upwind and downwind
pressure gradients have an almost equal contribution. Analogously, we compute
u
Ci+1,j as the weighted average:
u u u
Ci+1,j = W (−P/2)Ci+1/2,j + W (P/2)Ci+3/2,j ,
u u u
such that Ci+1,j is given by the average of Ci+1/2,j and Ci+3/2,j for |P | 1
and by the upwind approximation if |P | 1.
4.3 Closure of the model

In the previous section, we presented integral representations for approximat-
ing the cell-face velocity component ui+1,j . To compute the discrete convective
operator (N u (u))i+1/2,j we also need ui+1/2,j+1/2 and vi+1/2,j+1/2 . These ve-
locity components are also computed using local BVPs, as shown in Figure 4.5.
For computing both ui+1/2,j+1/2 and vi+1/2,j+1/2 , we make use of the homogen-
eous 1-D local BVP method discussed in Section 4.2. The velocity component
ui+1/2,j+1/2 is computed from the local boundary value problem
V uy − uyy = 0, yj < y < yj+1 , x = xi+1/2 , (4.22a)

u(xi+1/2 , yj ) = ui+1/2,j , u(xi+1/2 , yj+1 ) = ui+1/2,j+1 , (4.22b)
where V is an estimate for vi+1/2,j+1/2 . Likewise, for vi+1/2,j+1/2 , we use the

following local BVP :
U vx − vxx = 0, xi < x < xi+1 , y = yj+1/2 , (4.23a)

v(xi , yj+1/2 ) = vi,j+1/2 , v(xi+1 , yj+1/2 ) = vi+1,j+1/2 , (4.23b)
45
vi+1/2,j+1/2
ui+1/2,j+1
yj+1 vi,j+1/2 vi+1,j+1/2
yj+1/2 ui+1/2,j+1/2
∆y
yj
yj ui+1/2,j
xi xi+1/2 xi+1
xi ∆x
(a) Velocity component ui+1/2,j+1/2 . (b) Velocity component vi+1/2,j+1/2 .
Figure 4.5: Computation of the other cell-face velocity components.
with U being an estimate for ui+1/2,j+1/2 . The computation of ui+1/2,j+1/2

using BVP (4.22) requires an estimate of the interface velocity vi+1/2,j+1/2
(denoted by V in equation (4.22a)), and vice versa, for the computation of
vi+1/2,j+1/2 , an estimate of ui+1/2,j+1/2 (denoted by U in equation (4.23a)) is
needed. Hence, we couple the two local BVPs (4.22) and (4.23). Note that
equation (4.22a) is the linearized form of the equation
(uv)y − uyy = 0,
which is the cross-flux gradient in the x-momentum equation. Similarly, equa-

tion (4.23a) is the linearized form of,
(uv)x − vxx = 0,
which is the cross-flux gradient in the y-momentum equation.

All the methods described so far are analogously applied for the discret-
ization of the convective term (N v (u) i,j+1/2 given by equation (4.4). To

compute the discretized operator (N v (u) i,j+1/2 , we need the velocity com-
ponents vi,j+1 , ui+1/2,j+1/2 and vi+1/2,j+1/2 . The velocity component vi,j+1 is
computed using the local BVP
1 v
(vy − bv)y = (δy p + Ci,j+1 ), x = xi , yj+1/2 < y < yj+3/2 ,

v(xi , yj+1/2 ) = vi,j+1/2 , v(xi , yj+3/2 ) = vi,j+3/2 ,
where b := V /, with V being an estimate for vi,j+1 . Again, we assume that
v
δy p is piecewise constant and Ci,j+1 ≈ (uv − vx )x |(xi ,yj+1 ) , is constant over the
interval (yj+1/2 , yj+3/2 ). Solving the local BVP, as discussed in Section 4.2, we
46
get that vi,j+1 is given by
vi,j+1 = v h ( 12 ) + v p ( 12 ) + v c ( 21 ), (4.24a)
v v
v h ( 21 ) = W (−P /2)vi,j+1/2 + W (P /2)vi,j+3/2 , (4.24b)
1
v p ( 21 ) = − ∆x2 H2 (−P v /2)(δy p)i,j+1/2 + H2 (P v /2)(δy p)i,j+3/2 , (4.24c)
4
1 v
c 1
v ( 2 ) = − Ci,j+1 tanhc(P v /4), (4.24d)
8
where the Péclet number P v is defined as P v := V ∆y/. Analogous to Ci+1,j

u
,
v
the discrete cross-flux gradient Ci,j+1 is computed as the weighted average
v
Ci,j+1 = W (−P v /2)Ci,j+1/2
v
+ W (P v /2)Ci,j+3/2
v
.
4.4 Iterative computation of the cell-face

velocities
The discussion so far gave details about the derivation of an integral repres-
entation of the cell-face velocity components by solving linearized local BVPs.
However, the local BVPs that we intend to solve are non-linear. To account for
this non-linearity, the cell-face velocities are computed iteratively. Algorithm
4.1 gives the details of the iteration procedure.
For the iterative computation of ui+1,j , we begin by making an estimate
for U in the linearization of equation (4.5). Corresponding to this estimate
we obtain the Péclet number given by P = U ∆x/, which is used to compute
the velocity component using one of the three methods described in the pre-
vious section. The computed cell-face velocity component is used to update
the estimate U and the Péclet number and the procedure is repeated until
the computed values converge to a given tolerance. The velocity components
ui+1/2,j+1/2 , vi+1/2,j+1/2 and vi,j+1 are computed iteratively in a similar man-
ner.
4.5 Numerical results

We now compare the local BVP methods with the central and upwind schemes
for the following test cases: the Taylor-Green vortex problem and the standard
lid-driven square cavity flow.
4.5.1 Taylor-Green vortex

The Taylor-Green vortex in two dimensions admits an exact solution to the
47
incompressible Navier-Stokes equations, given by

2
u(x, y, t) = − sin(πx) cos(πy)e−2π t
, (4.25a)
2
−2π t
v(x, y, t) = cos(πx) sin(πy)e , (4.25b)
1 2
p(x, y, t) = (cos(2πx) + sin(2πy))e−4π t . (4.25c)
4
We compare the local BVP methods with the central and the upwind schemes
on the domain [ 41 , 2 14 ] × [ 14 , 2 14 ] with periodic boundary conditions. For time
integration, we use the second-order accurate, energy-conserving implicit Gauss
method (the midpoint rule, [75]) with a fixed time step (∆t = 10−2 ). The
simulation is run from t = 0 to t = 100, i.e., 10, 000 time steps are taken. Table
4.1 shows the discretization error (keu k∞ = kunum − uexact k∞ ) for Re = 105 on
a 20 × 20 grid, at times t = 20, 100. At t = 20, the numerical solution obtained
using the central scheme does not exhibit oscillations, and is significantly more
accurate than the solutions of the upwind and homogeneous 1-D local BVP
method. The central scheme and the inhomogeneous 1-D and 2-D local BVP
methods appear to produce practically identical results at t = 20. However, at
t = 100, the central scheme gives an oscillating solution, which severely affects
the accuracy. The spurious oscillations can be observed through the streamlines
of the numerical solutions, as shown in Figure 4.6. For two-dimensional flows
with smooth initial conditions, the incompressible Navier-Stokes equations have
a unique solution [51, 52]. Unlike the central scheme, the numerical solutions
obtained using upwind, homogeneous 1-D local BVP, inhomogeneous 1-D local
BVP and the 2-D local BVP scheme exhibit the same streamlines as the exact
solution (4.25).
The breakdown of the central scheme is shown by the dissipation of the
discrete kinetic energy. The central scheme results in a skew-symmetric con-
vective operator, which has the property that the net work done by convection
is zero. Further, for periodic boundary conditions it can be shown that work
done by the pressure is also zero, thus the dissipation of discrete kinetic energy
is only due to the discrete diffusion operator (for details see [97]). Moreover,
the implicit Runge-Kutta methods presented in [75] are shown to be energy-
conserving, thus for the central scheme the decay of the kinetic energy is solely
due to the discrete diffusion operator D. Figure 4.7 shows the dissipation of
the discrete kinetic energy for all the methods applied to the Taylor-Green vor-
tex flow. As expected, the upwind discretization shows a high dissipation of
kinetic energy. For a coarse grid and high Re (high Péclet numbers), the ho-
mogeneous 1-D local BVP method has a higher contribution from the upwind
part uup than from the central part ucen , cf. equation (4.20). Consequently,
the homogeneous 1-D local BVP method has similar dissipation of energy as
the upwind scheme. The inhomogeneous 1-D and 2-D local BVP methods ex-
hibit the same dissipation of the kinetic energy as the central scheme, until the
latter breaks down (at about t = 70). Computation of the discrete kinetic en-
ergy, requires approximating the velocity components at the cell centers, which
48
Algorithm 4.1: Iteration to compute the cell-face velocity component

ui+1,j .
Initialization: Set uh = 12 (ui+1/2,j + ui+3/2,j ), up = 0 and uc = 0
ui+1,j = uh + up + uc
Set TOL as allowed maximum tolerance for convergence

Set error > TOL
while error ≥ TOL do
uold
i+1,j = ui+1,j
P := 1 ui+1,j ∆x
Compute uh , up and uc from (4.14), (4.15) and (4.16), respectively
ui+1,j = uh + up + uc

error = ui+1,j − uold
i+1,j

end
Method keu k∞ (t = 20) keu k∞ (t = 100)
Central 2.9897 × 10−5 1.7822

Upwind 8.2059 × 10−1 8.8515 × 10−1
Homogeneous 1-D local BVP 8.2051 × 10−1 8.8512 × 10−1
Inhomogeneous 1-D local BVP 2.9897 × 10−5 1.5480 × 10−4
2-D local BVP 2.9897 × 10−5 1.5480 × 10−4
Table 4.1: Absolute errors (keu k∞ ) for the Taylor-Green vortex problem, for
Re = 105 , on a 20 × 20 grid, at t = 20 and t = 100.
49
Figure 4.6: Streamline plots for the Taylor-Green vortex problem for Re = 105
on a 20 × 20 grid at t = 100 (∆t = 10−2 , implicit second-order accurate Gauss
method)
50
Exact
Central
Upwind
E kin
0.5
Hom 1-D
Inhom 1-D
Inhom 2-D
0
0 50 100
t
Figure 4.7: Temporal behavior of the discrete kinetic energy for the Taylor-
Green vortex problem, for Re = 105 .
is done by averaging. The availability of the analytical solution allows us to

compute exact velocity components at the cell centers. Therefore, we observe
the difference between the kinetic energy of the exact solution (4.25) and the
numerical solutions. The inclusion of the source term in the local BVP ap-
pears to nullify the effect of numerical dissipation from the homogeneous part.
Although the central scheme results in a symmetry-preserving discretization,
which is energy-conserving [97], the presence of oscillations in its numerical
solution results in non-physical dissipation of energy, as observed in Figure 4.7.
Next, we investigate the convergence of spatial errors over a family of uni-
form grids ranging from 5 × 5 to 320 × 320 for Re = 100. For this case we use
the fourth-order, four-stage explicit Runge-Kutta method, just as in [76], to
ensure that for sufficiently small time step the temporal errors are negligible
compared to the spatial errors. Figure 4.8 shows the convergence of the errors
kunum − uexact k∞ for Re = 100 over a family of uniform grids, for a fixed
time step ∆t = 10−3 . It can be observed that the upwind discretization of the
convective terms results in a first-order accurate scheme. The homogeneous
1-D local BVP method has a dominant upwind (uup ) contribution for coarser
grids (due to higher Péclet numbers), hence exhibits first-order accuracy there.
However, for finer grids (decreasing Péclet numbers), the central part (ucen )
becomes dominant, leading to second-order convergence. The inclusion of the
source terms in the local BVP improves the accuracy of the computed numer-
51
10 0
Upwind
10 -2
keu k∞
Hom 1-D
2
10 -4
Central
Inhom 1-D
Inhom 2-D
10 -6
10 -3 10 -2 10 -1 10 0
h
Figure 4.8: Convergence of the spatial errors for the Taylor-Green vortex prob-
lem, for Re = 100 and at t = 20, for a family of grids (5 × (2i , 2i ), i =
0, 1, 2, · · · , 6 ).
ical solution significantly, resulting in the same error behavior (same order and
same magnitude) as that of the central scheme.
4.5.2 Lid-driven cavity flow

This internal flow problem is very well suited for studying the effect of including
the cross-flux term in the computation of cell-face velocity components due to
the strong two-dimensionality of the flow. We investigate the lid-driven cavity
flow for the case Re = 100, and use the results obtained by Ghia, Ghia and
Shin [29] as reference solutions. We apply a pair of fairly coarse grids, viz.,
8 × 8 and 16 × 16 for the Re = 100 case. Figures 4.9 and 4.10 show the plots of
the horizontal velocity component u along the vertical center line of the cavity
on the two grids. We compare the velocity profiles obtained using the central
scheme, the upwind scheme and the three methods described in Section 4.2.
The upwind method yields the least accurate results, as observed from Figures
4.9 and 4.10. For the 8 × 8 grid, we have higher Péclet numbers, thus the
homogeneous 1-D local BVP method behaves much like the upwind method.
The inclusion of the discrete pressure gradient in the inhomogeneous 1-D local
BVP method enhances the accuracy of the homogeneous 1-D local BVP method
and gives velocity profiles similar to the central scheme as observed in both
Figures 4.9 and 4.10. As anticipated, the 2-D local BVP method performs the
52
Method Cost ratio
Central 1
Upwind 0.911
Homogeneous 1-D local BVP 1.106
Inhomogeneous 1-D local BVP 1.263
2-D local BVP 1.284
Table 4.2: Comparison of computing times for the lid-driven cavity problem,
for Re = 100, on a uniform 16 × 16 grid.
best of the considered methods, due to the strong two-dimensionality of the

flow.
Table 4.2 shows the ratio of the computing times of the upwind method and
the three local BVP methods over the computing time of the central method,
for the 16 × 16 grid case. The costs of the local BVP methods appear to be
only marginally higher than the cost of the central method.
For coarse grids the inclusion of the source terms in the local BVP for the
approximation of the cell-face velocity components results in higher accuracy
than the central scheme. Further, we observe that the inclusion of the source
terms appears to cancel the dissipative effects arising from the upwind part of
the homogeneous part of the cell-face velocity component.
4.6 Conclusion
In this chapter we presented methods for computing the cell-face velocity com-
ponents involved in the discrete convective terms of the incompressible Navier-
Stokes equations, using local BVPs. The cell-face velocities are computed by
iteratively solving the nonlinear, local BVPs. Depending on the terms in the
right-hand side of the local BVP we get three different methods, namely, the ho-
mogeneous 1-D local BVP method, the inhomogeneous 1-D local BVP method
and the 2-D local BVP method. The homogeneous 1-D local BVP method
provides an estimate for the cell-face velocity as a weighted average of the neigh-
boring values depending on the Péclet number. In the limiting case P → 0, it
becomes identical to the central scheme and for |P | → ∞ it reduces to the up-
wind scheme. The order of accuracy of the computed numerical solutions using
the homogeneous 1-D local BVP method varies from first-order to second-order
depending on the mesh Péclet numbers.
In the inhomogeneous 1-D local BVP method we include the effect of the
53
Ghia-Ghia-Shin Upwind Inhomogeneous 1-D local BVP

Central Homogeneous 1-D local BVP 2D local BVP
1.0
0.8
0.8
0.6
0.6
y
0.4
0.4
0.2
0.2
0.0
−0.4 −0.2 0.0 0.2 0.4 0.6 0.8 1.0
u
Figure 4.9: Horizontal velocity component along the vertical center line of the
cavity for Re = 100 on a uniform 8 × 8 grid.
pressure gradient, where it is assumed that the pressure is piecewise linear.

As an enhancement of the inhomogeneous 1-D local BVP method we have the
2-D local BVP method where we include both the pressure gradient and the
cross-flux gradient. Unlike the homogeneous 1-D local BVP method, the inclu-
sion of the source terms results in second-order accurate numerical solutions
for all values of P considered. As observed for the case of the Taylor-Green
problem, both the inhomogeneous 1-D local BVP method and the 2-D local
BVP method exhibit perfect kinetic energy dissipation. The central scheme,
which suffers from spurious oscillations, fails at higher Péclet numbers. The
local BVP methods do not exhibit oscillations, which can be credited to the
presence of the upwind component in the homogeneous part of the cell-face
velocities.
For incompressible flows, the numerical scheme is energy conserving (in
the inviscid case) if the discrete convective operator is skew-symmetric, which
happens if the cell-face velocities are computed by central averaging. For the
local BVP methods, the discrete convective operator is not skew-symmetric,
thus the work done by the convective forces is formally not zero. Nevertheless,
due to the source terms, the inhomogeneous 1-D local BVP method and the
2-D local BVP method do not show any dissipative behavior.
Of the three local BVP methods, the 2-D local BVP method is to be pre-
ferred as it results in a more accurate approximation at marginally higher cost.
54
Ghia-Ghia-Shin Upwind Inhomogeneous 1-D local BVP

Central Homogeneous 1-D local BVP 2-D local BVP
1.0
0.8
0.8
0.6
0.6
y
0.4
0.4
0.2
0.2
0.0
−0.4 −0.2 0.0 0.2 0.4 0.6 0.8 1.0
u
Figure 4.10: Horizontal velocity component along the vertical center line of the
cavity for Re = 100 on a uniform 16 × 16 grid.
55
Chapter 5
Complete-flux schemes
In this chapter we first rehearse the complete-flux scheme (CFS) for the linear
convection-diffusion-reaction equation as presented in [87]. Analogous to the
local BVP approximation schemes presented in Chapter 4, in the CFS we first
derive an integral representation for the total flux from a local two-point BVP.
The integral representation is derived from a local BVP for the entire equation,
i.e., including the source term and therefore, it can be expressed as a sum of
a homogeneous and an inhomogeneous part, corresponding to the convection-
diffusion operator and the source term, respectively. The numerical flux is then
computed from the integral representation using suitable quadrature rules. The
CFS scheme is then extended to the viscous Burgers equation.
For the Burgers equation, the nonlinearity of the flux makes it difficult
to derive a numerical flux expression from the integral representation that is
consistent in the inviscid limit. Moreover, we show that using straightforward
quadrature rules for the nonlinear integral formulation of the flux does not
result in a consistent approximation. To simplify the flux computation, we
linearize the local BVP using the cell-face velocities, which are computed using
local BVPs, as demonstrated in the previous chapter. The linearized BVP is
then used to compute the flux. For the linearization, the choice of the method
used to compute the cell-face velocity plays a significant role, in determining
the flux in the vanishing viscosity limit.
This chapter is organized as follows: Section 5.1 presents the derivation of
the integral representation of the fluxes for the one-dimensional inhomogen-
eous convection-diffusion equation using local BVPs. In Section 5.2 we extend
the scheme to the viscous Burgers equation using linearized local BVPs. Sec-
tion 5.3 presents the approximation of the flux using nonlinear local BVPs.
The behavior of both the linear and the nonlinear scheme in the invscid limit
is presented in Section 5.4 and conclusions are drawn in Section 5.5.
57
5.1 Integral representation of the flux

To illustrate the computation of the numerical flux from a local BVP, we con-
sider the one-dimensional inhomogeneous convection-diffusion equation
φt + fx = s, f := aφ − νφx , (5.1)
with constant diffusion coefficient ν, velocity field a = a(x) and source term
s = s(x, t). The flux fi+1/2 at the cell-face xi+1/2 = (xi + xi+1 )/2 is computed
from the local BVP
fx = (aφ − νφx )x = s, xi < x < xi+1 , (5.2a)

φ(xi ) = φi , φ(xi+1 ) = φi+1 . (5.2b)
Note that in the above BVP we ignore the time-dependence of the flux. For
unsteady problems, the time derivative can be included as part of the source, as
demonstrated in Chapter 3 for the convective discretization of the convection-
diffusion equation. For the following derivation we define the variables
Z x Z x
a ah
λ := , P := , Λ(x) := λ(ξ)dξ, S(x) := s(ξ)dξ,
ν ν xi+1/2 xi+1/2
with h = xi+1 − xi and Λ(x) the Péclet integral. Integrating (5.2a) from xi+1/2
to x ∈ (xi , xi+1 ) gives
f (x) − fi+1/2 = S(x).
Substituting the integrating factor formulation f = −ν φe−Λ(x) x eΛ(x) in (5.2a),

and integrating from xi to xi+1 and applying the boundary conditions gives
the integral representation of the flux, i.e.,
h inh
fi+1/2 =fi+1/2 + fi+1/2 , (5.3a)
. Z xi+1
−Λi −Λi+1
h
fi+1/2 =ν e φi − e φi+1 e−Λ dx, Λi := Λ(xi ), (5.3b)
xi
Z xi+1 . Z xi+1
−Λ
inh
fi+1/2 = − e Sdx e−Λ dx, (5.3c)
xi xi
h inh
with fi+1/2 and fi+1/2 corresponding to the homogeneous and the particular
h inh
solution of the BVP, respectively. Consequently, we refer to fi+1/2 and fi+1/2
as the homogeneous and inhomogeneous component of the flux. Note that
the integral representation provides an exact solution of the local BVP (5.2).
Consequently, the numerical fluxes are locally exact. To simplify the integral
representations (5.3) we introduce hu, vi as the inner product of two functions
u = u(x) and v = v(x) defined on (xi , xi+1 ), i.e.,
Z xi+1
hu, vi = u(x)v(x)dx.
xi
58
Using the inner product, the integral representation of the flux (5.3) can be
formulated as
h e−Λi φi − e−Λi+1 φi+1

fi+1/2 =ν , (5.4a)
he−Λ , 1i
inh he−Λ , Si
fi+1/2 =− . (5.4b)
he−Λ , 1i
Constant coefficients. For problems with constant flow speed a and a con-
stant s = si+1/2 , all integrals in (5.4) can be evaluated exactly. In this case,
we have λ = a/ν constant over (xi , xi+1 ), giving
hλ, 1i = λh = P.
Further, the Péclet integral and the source integral are given by
Λ(x) = λ(x − xi+1/2 ), S(x) = si+1/2 (x − xi+1/2 ).
For the constant coefficient case, the numerical flux Fi+1/2 is then given by
h inh
Fi+1/2 =Fi+1/2 + Fi+1/2 , (5.5a)
h ν
Fi+1/2 = B(−P )φi − B(P )φi+1 , (5.5b)
h
1
inh
Fi+1/2 =h − W1 (P ) si+1/2 , (5.5c)
2
where the functions B and W1 are defined as
z ez − 1 − z
B(z) := , W1 (z) := ,
ez −1 z(ez − 1)
and depicted in Figure 5.1. Note that B(z) is the generating function for the
Bernoulli numbers. We refer to it as the Bernoulli function. The function
W1 (z) satisfies W1 (z) + W1 (−z) = 1, with 0 ≤ W1 (z) ≤ 1 and W1 (0) = 1/2.
inh
Thus, for P = 0, we get fi+1/2 = 0 and the source term has no effect on the
flux.
h
The homogeneous component Fi+1/2 can also be expressed as a weighted
sum of the central and the upwind flux, i.e.,
h cen
up
Fi+1/2 = 2W1 (|P |)Fi+1/2 + 1 − 2W1 (|P |) Fi+1/2 ,
cen a ν
Fi+1/2 = (φi + φi+1 ) − (φi+1 − φi ),
2 h
up up ν
Fi+1/2 = a φi+1/2 − (φi+1 − φi ),
h
cen up
where Fi+1/2 and Fi+1/2 are the first-order upwind and the second-order central
approximation of the flux function. Figure 5.2 (a) shows the weight function
59
10 1
W1 (z)
B(z)
5 0.5
0 0
-10 0 10 -10 0 10
z z
Figure 5.1: The Bernoulli function B(z) and the function W1 (z).
2W1 (|P |). It can be seen that for |P | 1, the homogeneous component reduces
to the central approximation and for |P | 1, to the upwind approximation.
inh
In the above derivation of the inhomogeneous component Fi+1/2 , the source
term s = si+1/2 is taken to be constant on (xi , xi+1 ). Alternatively, there may
be instances where the source is defined to be piecewise constant, such that
(
si , xi < x < xi+1/2 ,
s(x) = (5.7)
si+1 , xi+1/2 ≤ x < xi+1 .
One such instance can be found in Chapter 6, where due to the staggered grid
arrangement, the source term (the pressure gradient) needs to be piecewise
constant. Integrating the term he−Λ , Si with s being piecewise constant results
inh
in the inhomogeneous flux component Fi+1/2
inh

Fi+1/2 = h C(−P )si − C(P )si+1 , (5.8)
where
ez/2 − 1 − z/2
C(z) = .
z(ez − 1)
Figure 5.2 (b) shows the function C(z), with C(z) → 1/8 as z → 0 and C(z) → 0,
C(−z) → 1/2 as z → ∞. Moreover, the functions C(z) and W1 (z) satisfy
1
C(−z) − C(z) = − W1 (z).
2
Thus, for si = si+1 , the inhomogeneous part of the flux given by (5.8) reduces
to expression (5.5c), for constant source term.
60
(a) (b)
1 0.5
C(z)
C(−z)
2W1 (|z|)
0.5 0.25
0 0
-20 -10 0 10 20 -20 -10 0 10 20
z z
Figure 5.2: Plots of functions: (a) 2W1 (|z|) and (b) C(z).
Example 5.1. In Chapter 3 we gave the spatial discretization of the convection-

diffusion equation, with the discrete convective term given by the local BVP
approximation and the discrete diffusive term approximated by the second-
order central difference approximation. In Section 3.3 we showed that the
homogeneous approximation of the convective term introduces numerical dif-
fusion, which is given by
num P
νconv =ν tanh(P/4).
2
Similarly, for the homogeneous flux approximation, it can be shown that the
discretization is dissipative and introduces numerical diffusion. In [88], the
authors have shown that for the homogeneous flux scheme (HFS) the numerical
diffusion is given by
num
= ν P 21 − W1 (P ) .

νHFS
Figure 5.3 shows the relative numerical diffusion ν num /ν for the homogeneous
flux scheme and the homogeneous convective flux scheme. Clearly, the ho-
mogeneous flux approximation has lower numerical diffusion compared to the
homogeneous convective flux approximation. Thus using local BVPs for the
approximation of the net flux, i.e., the sum of the convective and the diffusive
flux, lowers the numerical diffusion, as compared to the homogeneous approx-
imation for the convective discretization.
The above example demonstrates that combining the convection and dif-
fusion operator, as in the homogeneous flux scheme, effectively reduces the
numerical dissipation introduced by the homogeneous convective flux approx-
imation, as introduced in Chapter 3. Using the flux expressions (5.5), we get
61
(a) (b)
10 1
num
νconv /ν
num
νHFS /ν
5 0.5
num num
(νconv − νHFS )/ν
0 0
0 10 20 0 10 20
P P
Figure 5.3: Comparison of numerical diffusion: (a) relative numerical diffusion

num num
as a function of the Péclet number, (b) The difference (νconv − νHFS )/ν as
function of the Péclet number.
that the numerical flux Fi+1/2 can be expressed as
Fi+1/2 = αi+1/2 φi + βi+1/2 φi+1 + γi+1/2 ŝi + δi+1/2 ŝi+1 ,
where ŝ = s − φ̇, αi+1/2 , βi+1/2 , γi+1/2 and δi+1/2 are coefficients depending
on ν, h and P . Thus the discretization results in the linear ODE system
BΦ̇ + AΦ = Bs + b, (5.9)
where Φ and s are vectors of the unknown and the source term, and where b
contains the boundary data. The tridiagonal matrices A and B correspond to
the homogeneous and inhomogeneous discretization, respectively.
In order to have a second-order convergent numerical method, we make

use of the θ-method (see Chapter 2) for time integration of the discrete sys-
tem (5.9), with θ = 1/2. An extensive collection of numerical results for
the complete-flux scheme for the convection-diffusion-reaction equation can be
found in [87, 101]. Further, convergence analysis of the complete-flux scheme
on unstructured grids can be found in [24]. For two-dimensional flows, a modi-
fication of the complete-flux scheme to locally flow adapted coordinates can be
found in [85]. We now extend the flux approximation to the viscous Burgers
equation.
62
5.2 Complete-flux scheme for the viscous

Burgers equation
In the conservative form the one-dimensional viscous Burgers equation reads

ut + 12 u2 − νux = 0.
x
For the above equation, the nonlinear flux function is given by
f = 12 u2 − νux .
The finite-volume discretization of the above equation results in the semi-

discrete system
h u̇i + Fi+1/2 − Fi−1/2 = 0,
where Fi+1/2 is the numerical approximation of the flux function at the interface
xi+1/2 . The numerical flux Fi+1/2 is computed from the local two-point BVP

fx = 21 u2 − νux = s, x ∈ (xi , xi+1 ), (5.10a)
x
u(xi ) = ui , u(xi+1 ) = ui+1 , (5.10b)
where the discrete time derivative can be treated as a source term, i.e.,
s = −u̇i+1/2 .
The nonlinearity of equation (5.10a) makes it difficult to derive explicit ex-
pressions for the numerical flux. In order to simplify the problem, we linearize
(5.10a) using the cell-face velocity. Let U = ui+1/2 ≈ u(xi+1/2 ) be an approx-
imation of the unknown u at the interface, then the flux can be computed from
the linearized BVP

1
2 U u − νux = s, x ∈ (xi , xi+1 ), (5.11a)
x
u(xi ) = ui , u(xi+1 ) = ui+1 . (5.11b)
For U and ν constant over [xi , xi+1 ], the problem is similar to the constant
coefficient convection-diffusion equation. Thus, using relations (5.5)-(5.8) we
lin
get that the (linearized) flux Fi+1/2 is given by
lin lin,h lin,inh

Fi+1/2 =Fi+1/2 + Fi+1/2 , (5.12a)
lin,h ν
Fi+1/2 = B(−Pi+1/2 ) ui − B(Pi+1/2 ) ui+1 , (5.12b)
h
1
lin,inh
Fi+1/2 =h − W1 (Pi+1/2 ) si+1/2 , (s is constant) (5.12c)
2
lin,inh
Fi+1/2 =h C(−Pi+1/2 )si − C(Pi+1/2 )si+1 , (s is piecewise constant),
(5.12d)
63
with Pi+1/2 = U h/2ν and expressions (5.12c) and (5.12d) corresponding to

constant and piecewise constant source term on (xj , xj+1 ), respectively. For
the above discretization, the linearization factor, i.e., the cell-face velocity,
ui+1/2 is computed using the homogeneous 1-D local BVP scheme, described
in Chapter 4. In the following example we demonstrate the effects of the cell-
face velocity approximation on the homogeneous flux approximation F lin,h .
Example 5.2. To compare the effects of the approximation methods for the
cell-face velocity used for the linearization of the local BVP, we examine the
lin lin,h
homogeneous flux scheme, i.e., Fi+1/2 = Fi+1/2 . Further, we use the standard
explicit four-stage Runge-Kutta method for the temporal discretization with
fixed time step ∆t = 10−4 . The 1-D Burgers equation defined on (x, t) ∈
(0, 1) × (0, T ) has an analytic solution (see [103], Chapter 4) given by:
r1 r2
uref (x, t) = 1 − 0.9 − 0.5 ,
r1 + r2 + r3 r1 + r2 + r3
where the terms r1 , r2 and r3 are given by
x − 0.5 99t x − 0.5 3t x− 3
2
r1 = exp − − , r2 = exp − − , r3 = exp − .
20ν 400ν 4ν 16ν 2ν
Using the analytical solution as a reference solution, we compare the perform-
ance of the homogeneous flux scheme for ν = 10−2 , 10−3 . Figures 5.4 and
5.5 show the convergence of the absolute errors keu k1 = ku − uref k1 over a
family of uniform grids (h = 2−i × 10−1 , i = 0, 1, · · · , 5) for ν = 10−2 and
ν = 10−3 , respectively. Although in all three methods the numerical flux
is given by the solution of the homogeneous linearized local BVP (5.12b), the
only difference is in the approximation method used for computing the cell-
face velocity, and thus the Péclet number Pi+1/2 . Since the upwind scheme
results in a first-order accurate approximation, the resulting homogeneous flux
approximation is also first-order consistent. The central scheme provides a
second-order accurate approximation for the cell-face velocity but introduces
spurious oscillations, particularly on coarse grids. For both ν = 10−2 , 10−3 ,
the homogeneous local BVP approximation of the cell-face velocity results in
a more accurate discretization than the central or upwind scheme (for cell-face
velocity).
In order to understand the monotonicity preserving property of the homo-
geneous flux scheme, we look at the total variation of the scheme.
Definition 5.3. Let X

TV(un ) := |uni − uni−1 |
i
be the total variation of the discrete solution at time level tn = n ∆t. A nu-
merical method is said to be total variation diminishing (TVD) if TV(un+1 ) <
TV(un ) or total variation non-increasing (TVNI) if TV(un+1 ) ≤ TV(un ).
64
10 -1
Central
Homogeneous local BVP
Upwind
10 -2 1
keu k1
10 -3
2
10 -4
10 -3 10 -2 10 -1
h
Figure 5.4: Convergence of the 1-norm of the absolute error for ν = 10−2 over
a family of uniform grids with h = 2−i × 10−1 , i = 0, 1 · · · , 5 for the linearized
homogeneous flux scheme, combined with the explicit four-stage Runge-Kutta
method (∆t = 10−4 ).
TVD schemes remove any numerical oscillation arising due to the spatial or
temporal discretization [40]. Figure 5.6 shows the total variation of the numer-
ical solutions computed using the linearized homogeneous flux approximation
with central, upwind and the homogeneous local BVP methods for the approx-
imation of the cell-face velocity as in Example 5.2, and the standard explicit
four-stage Runge-Kutta method. Both the upwind and the homogeneous local
BVP method do not introduce any spurious oscillations; they both result in
a TVNI discretization. The central scheme, due to the relatively high Péclet
numbers (max{|P |} = 50), suffers from spurious oscillations which can be seen
from the change of the total variation.
In the following, the cell-face velocity for the linearized homogeneous flux
approximation is computed using the homogeneous convective flux approxima-
tion (described in Chapter 4), such that the flux approximation is always free of
any spurious oscillations. However, the homogeneous flux approximation being
the weighted average of the central and the upwind scheme is not uniformly
second-order accurate. In order to have a uniformly second-order accurate flux
approximation, it is essential to include the inhomogeneous part of the flux. In
the following section, we demonstrate the effect of the source term on the flux
approximation.
65
10 -1
Central
Homogeneous local BVP
1
Upwind
10 -2
keu k1
10 -3
10 -4
10 -3 10 -2 10 -1
h
Figure 5.5: Convergence of the 1-norm of the absolute error for ν = 10−3 over
a family of uniform grids with h = 2−i × 10−1 , i = 0, 1 · · · , 5 for the linearized
homogeneous flux scheme combined with the explicit four-stage Runge-Kutta
method (∆t = 10−4 ).
5.2.1 Inhomogeneous flux approximation

In the linearized local BVP (5.11) we included the time derivative of the un-
known as a source term. The assumption of the source term being either
constant or piecewise constant on the interval (xi , xi+1 ) reduces the inhomo-
geneous part of the flux to (5.12c) or (5.12d), respectively. Inclusion of the
inhomogeneous part of the flux changes the semi-discretization to the implicit
ODE-system
(Ω + B U̇ + AU = b, (5.13)
where the matrix B corresponds to the inhomogeneous part of the flux ap-
proximation. Compared to the linear system (5.9), the above system has a
modified mass matrix, Ω + B. For time integration of the above system we
use the θ-method with θ = 1/2. With periodic boundary condition the above
system can be expressed as
LUn+1 = RUn .
For the above numerical scheme to be TVD, it is necessary for the operator L
to be total variation increasing (TVI) and for R to be TVNI, i.e.,
TV(Lv) ≥ TV(v), TV(Rv) ≤ TV(v).
Then the numerical scheme is TVD [40]. For the system (5.13), a restriction
on the time step can be derived such that the discretization is always TVD.
66
1.5
TV →
Upwind
Homogeneous
Central
0 0.5 1
t→
Figure 5.6: The total variation for the linearized homogeneous flux scheme
with different methods for computing the cell-face velocities. Parameter values:
ν = 10−3 , h = 5 × 10−2 and ∆t = 10−4 .
Figure 5.7 and Figure 5.8 show the convergence of the errors for the lin-
earized inhomogeneous flux approximations compared to the homogeneous flux
approximation for ν = 10−2 and ν = 10−3 , respectively. It can be observed that
for very coarse grids (max(|P |) 1) the inclusion of the source term does not
appear to improve the accuracy of the homogeneous approximation, particu-
larly for ν = 10−3 (Figure 5.8), where the solution u(x, t) of the viscous Burgers
equation has a steep gradient, which causes the numerical solution, computed
using the inhomogeneous flux approximation, to exhibit an overshoot or under-
shoot. For coarse grids it is observed (for both ν = 10−2 and ν = 10−3 ) that a
constant source term results in a more accurate approximation than a piecewise
constant source term. On finer grids, the piecewise constant source term res-
ults in a more accurate flux approximation. For ν = 10−2 , the inhomogeneous
approximation (with piecewise constant source) results in a uniformly second-
order accurate approximation. For ν = 10−3 the inhomogeneous approximation
does exhibit second-order accuracy over fine grids, as undershoots/overshoots
may occur (on coarse grids) in regions with steep gradient.
It is recommended that the source term is assumed to be constant (if pos-

sible) over coarse grids (|P | > 10) and piecewise constant over finer grids.
Overshoots and undershoots can be avoided by using appropriate flux limiters,
as demonstrated in [101] for the linear convection-diffusion equation. To com-
pletely remove the spurious extrema, TVD conditions should be imposed such
that L is TVI and R is TVNI.
67
Figure 5.7: Convergence of the 1-norm of the absolute error for ν = 10−2
for a family of uniform grids with h = 2−i × 10−1 , i = 1, · · · , 5, for the
linearized homogeneous and inhomogeneous flux schemes, combined with the
semi-implicit θ-method with θ = 1/2 and ∆t = 10−3 .
Figure 5.8: Convergence of the 1-norm of the absolute error for ν = 10−3
for a family of uniform grids with h = 2−i × 10−1 , i = 0, · · · , 5, for the
linearized homogeneous and inhomogeneous flux schemes, combined with the
semi-implicit θ-method with θ = 1/2 and ∆t = 10−3 .
68
5.3 Nonlinear local BVP approximation

So far, we have computed the numerical flux from the linearized local BVP
(5.11), giving us an integral representation for the linearized flux function
1
f lin = U u − νux .
2
In order to derive the fully nonlinear approximation for the flux function
1 2
f= u − νux ,
2
we need to solve the nonlinear BVP (5.10). Using the integral representations
(5.4) for the convection-diffusion equation, with convective term a = u(x)/2
(not constant), we can derive the integral flux representation for the viscous
Burgers equation. For the nonlinear local BVP
1
u2 − νux = −ut , xi < x < xi+1 , (5.14a)
2 x
u(xi ) = ui , u(xi+1 ) = ui+1 , (5.14b)
the flux is given by the implicit integral representation

h inh
f =fi+1/2 + fi+1/2 , (5.15a)
−Λi −Λi+1
h e ui − e ui+1
fi+1/2 =ν , (5.15b)
he−Λ , 1i
inh he−Λ , Si
fi+1/2 =− , (5.15c)
he−Λ , 1i
with
Z x Z x
λ(x) = u(x)/2ν, Λ(x) = λ(ξ)dξ and S(x) = s(ξ)dξ.
xi+1/2 xi+1/2
Rewriting the Péclet integrals Λi and Λi+1 as

1
Λi = Λ̄i+1/2 − hλ, 1i,
2
1
Λi+1 = Λ̄i+1/2 + hλ, 1i,
2
with Λ̄i+1/2 = (Λi + Λi+1 )/2 (6= Λ(xi+1/2 )), the integral representation of the
h
homogeneous flux fi+1/2 can be expressed as
h ehλ,1i/2 ui − e−hλ,1i/2 ui+1

fi+1/2 =ν . (5.16)
heΛ̄i+1/2 −Λ , 1i
69
Taylor expanding the Péclet integral Λ(x) about xi+1/2 , i.e.,
1 (x − xi+1/2 )2
Λ(x) = (x − xi+1/2 )u(xi+1/2 ) + ux (xi+1/2 )+
2ν 4ν
(x − xi+1/2 )3
uxx (xi+1/2 ) + · · · ,
12ν
we conclude that the integrals Λi and Λi+1 are given by
h h2
Λi = − u(xi+1/2 ) + ux (xi+1/2 ) + O(h3 ), (5.17a)
4ν 16ν
h h2
Λi+1 = u(xi+1/2 ) + ux (xi+1/2 ) + O(h3 ), (5.17b)
4ν 16ν
which gives us that
h2
Λ̄i+1/2 = ux (xi+1/2 ) + O(h4 ). (5.18)
16ν
To derive the numerical flux from the integral representation (5.16), a
quadrature rule must be chosen such that the resulting discretization has a
3-point coupling. The integrals in (5.16) and (5.15c) could be evaluated us-
ing the trapezoidal rule or Newton-Cotes formulas, like Simpson’s rule and
the mid-point rule. Hence, for the homogeneous part of the flux, the integral
heΛ̄i+1/2 −Λ , 1i could be approximated as
h hλ,1i/2
heΛ̄i+1/2 −Λ , 1i ≈ + 4eΛ̄i+1/2 + e−hλ,1i/2 ,

Simpson : e (5.19a)
6
h
Mid-point : heΛ̄i+1/2 −Λ , 1i ≈ eΛ̄i+1/2 , (5.19b)
2
h
heΛ̄i+1/2 −Λ , 1i ≈ ehλ,1i/2 + e−hλ,1i/2 .

Trapezoidal : (5.19c)
2
Simpson’s rule and the mid-point approximation involve eΛ̄i+1/2 . From relation
(5.18), it can be observed that in the limit ν → 0, Λ̄i+1/2 may blow up. This
can be avoided by using the trapezoidal approximation. Therefore, we use the
trapezoidal rule to derive the numerical flux from the integral representation
(5.15).
The trapezoidal approximation. To derive the numerical approximation of

the homogeneous part of the flux, we need to evaluate the integrals hλ, 1i and
heΛ̄i+1/2 −Λ , 1i. In Example 5.2 we showed that the linearized homogeneous flux
scheme combined with the homogeneous approximation for the cell-face velocity
provides better accuracy than using the central approximation (equivalent with
the trapezoidal approximation of hλ, 1i) for the cell-face velocities. The integral
70
hλ, 1i is approximated using the mid-point rule, which results in the second-
order accurate approximation
ui+1/2 h
hλ, 1i ≈ = Pi+1/2 .
2ν
Using the mid-point approximation of hλ, 1i in the trapezoidal approximation
(5.19c) gives
h hλ,1i/2
heΛ̄i+1/2 −Λ , 1i ≈ + e−hλ,1i/2

e
2
h
≈ exp Pi+1/2 /2 + exp − Pi+1/2 /2 .
2
h
From the above, we get that the numerical approximation Fi+1/2 of the homo-
h
geneous part fi+1/2 is given by
h 2ν
Fi+1/2 = W (−Pi+1/2 )ui − W (Pi+1/2 )ui+1 , (5.20)
h
where W (z) = (ez + 1)−1 , is the weight function as defined in Chapter 4.
Similarly, using the trapezoidal rule, the inhomogeneous part of the integral
representation (5.15c) of the flux can be evaluated as
inh e−Λi Si + e−Λi+1 Si+1

Fi+1/2 ≈
e−Λi + e−Λi+1
= W (−hλ, 1i) Si + W (hλ, 1i) Si+1
≈ W (−Pi+1/2 ) Si + W (Pi+1/2 ) Si+1 .
As before, the source integrals Si and Si+1 depend on the approximation of
the source term, i.e., on whether the source term s = −u̇i+1/2 is assumed to
be constant or piecewise constant on (xi , xi+1 ). For constant source, u̇i+1/2 is
given by the upwind approximation of u̇ and the inhomogeneous part of the
flux reads
h
inh
= tanh 21 Pi+1/2 u̇up

Fi+1/2 i+1/2 . (5.21)
2
For piecewise constant source term we get
inh 1
Fi+1/2 = h W (−Pi+1/2 )u̇i − W (Pi+1/2 )u̇i+1 . (5.22)
2
In general, the numerical flux function resulting from a two-point approxim-
ation can be expressed as Fi+1/2 = F (ui , ui+1 ). The flux approximation is said
to be consistent if F (u, u) = f (u), where f (u) is the continuous flux function.
For a constant velocity u, the flux function for the viscous Burgers equation
is given by f (u) = u2 /2. For the linearized homogeneous flux approximation
(5.12b) we have
lin,h ν ν u2 uh
Fi+1/2 (u, u) = B(−P ) − B(P ) u = uP = , P = .
h h 2 2ν
71
Similarly, for the nonlinear homogeneous flux approximation (5.20) we get
h 2ν u2
Fi+1/2 = W (−P ) − W (P ) u = tanhc(P/2).
h 2
Thus, we get that for constant velocity u (6= 0) the nonlinear homogeneous
flux is smaller than the actual flux, i.e., u2 /2. This is also demonstrated when
we compare the nonlinear homogeneous flux approximation with the linearized
homogeneous flux F lin,h in the following example.
Example 5.4. We compare the convergence of the spatial discretization error
for the linear and the nonlinear homogeneous flux approximation schemes for
the 1-D viscous Burgers equation with ν = 10−2 on (x, t) ∈ (0, 1) × (0, T ).
Figure 5.9 shows the convergence of the errors on a family of uniforms grids with
h = 2−i × 10−1 , i = 0, 1, · · · , 5. Although the nonlinear scheme does exhibit
second-order convergence for finer grids, it is clearly less accurate than the
(linearized) homogeneous flux scheme. This is more clearly depicted through
the solution profiles for both schemes on a relatively coarse grid with h =
5 × 10−2 (Pmax = 5), as shown in Figure 5.10. The nonlinear approximation
appears to be more dissipative than the linear scheme, which can be attributed
to the breakdown of the scheme for smaller values of ν, as shown in the following
section.
5.4 Flux in the inviscid limit

We investigate the behavior of both the linear and nonlinear homogeneous flux
approximation in the limit ν → 0, i.e., P → ∞. For the following analysis, we
set uL := ui and uR := ui+1 .
Linearized homogeneous flux. The linearized homogeneous approximation
reads
ν U B(−P ) B(P )
F lin,h = B(−P )uL − B(P )uR = uL − uR .
h 2 P P
If U > 0, then in the inviscid limit P → ∞ the flux is given by
1
F lin,h = U uL .
2
Otherwise, for U < 0, P → −∞ and the limit flux is given by
1
F lin,h = U uR .
2
The cell-face velocity U , which is computed using the homogeneous local BVP
approximation, reduces to the upwind approximation in the inviscid limit, i.e.,
(
uL for s ≥ 0,
U=
uR for s < 0,
72
10 -1
Nonlinear homogeneous appr. 1
Linearized homogeneous appr.
10 -2
keu k1
10 -3 2
10 -4
10 -3 10 -2 10 -1
h
Figure 5.9: Convergence of the 1-norm of the error for ν = 10−2 on a family of
uniform grids with h = 2−i × 10−1 , i = 0, 1, · · · , 5, for the nonlinear homogen-
eous flux approximation (5.20) and the linearized homogeneous flux approx-
imation (5.12b) with the standard explicit four-stage Runge-Kutta method for
time integration (∆t = 10−4 ).
1
Reference
Linearized homogeneous appr.
Nonlinear homogeneous appr.
u→
0.5
0 0.5 1
x→
Figure 5.10: Solution for the linear and nonlinear homogeneous flux approx-
imation for the viscous Burgers equation for ν = 10−2 , h = 5 × 10−2 and
∆t = 10−4 (RK4).
73
where s = (uL +uR )/2 and F lin reduces to Roe’s approximation of the Riemann
problem [69, 70], i.e.,
(
1 2
Roe u for s ≥ 0,
F = 21 L 2
2 uR for s < 0.
Both, the linearized homogeneous flux approximation and Roe’s approximation

involve linearization of the convective term, and therefore do not provide the
correct flux for the nonlinear (Burgers equation) problem when uL ≤ 0 ≤ uR .
For this case, the nonlinear local BVP possesses two boundary layers for ν 1
(more details can be found in Section 7.3). The linearized homogeneous flux
approximation is only able to resolve just one of the two boundary layers and
hence, does not provide the correct flux approximation for uL ≤ 0 ≤ uR and
ν 1.
Nonlinear homogeneous flux. Observe that the homogeneous part (5.20)
can also be expressed as
W (−P ) W (P )
Fh = u uL − uR ,
P P
with P ≈ hλ, 1i, where λ(x) = u(x)/2ν. Now in the inviscid limit ν → 0, i.e.,
|P | → ∞, the limit flux is given by
F h = 0.
However, the limit flux is given by u2 /2. To understand the failure of the
homogeneous flux function we go back to the integral representation (5.15).
Note that the integral representation (5.15) provides an exact solution of the
local BVP (5.14). We now introduce the coefficients C1h and C2h , such that the
homogeneous part of the integral representation (5.15b) can be expressed as
h
= ν C1h uL − C2h uR ,

fi+1/2 (5.23a)
x x
he−Λ , 1i
Z Z
1 i+1
h
= −Λ
= heΛi −Λ , 1i = exp − 2ν 1
u(ξ)dξ dx, (5.23b)
C1 e i
xi xi
Z xi+1 Z xi+1
1 he−Λ , 1i Λi+1 −Λ 1

h
= −Λ
= he , 1i = exp 2ν u(ξ)dξ dx. (5.23c)
C2 e i+1 xi x
For the flux approximation to be consistent in the limit ν → 0 (with the

Godunov approximation for the inviscid Burgers equation), it is essential that
the integral coefficients reduce to
 
1
u , s ≥ 0, 0, s ≥ 0,
2 L

 


h h
νC1 = 0, s < 0, νC2 = − 21 uR , s < 0,

 

0, uL ≤ 0 ≤ uR , 0, uL ≤ 0 ≤ uR .
 
74
None of the quadrature rules considered in Section 5.3 results in an approxim-

ation that satisfies the above requirements.
5.5 Conclusion
In this chapter, we presented the approximation of the net flux, i.e., the sum of
the convective and the viscous flux using local BVPs. We showed that for the
constant coefficient convection-diffusion equation, the homogeneous local BVP
approximation of the flux is less dissipative than the discretization presented
in Chapter 3, where the convective term is approximated using a homogeneous
local BVP and the diffusion term by the second-order central approximation.
The flux approximation scheme is then extended to the viscous Burgers
equation which has a nonlinear flux function. To simplify the computation the
flux function is linearized using the cell-face velocity which is computed using
the local BVP, as shown in Chapter 4. It is shown that using the homogeneous
local BVP method for the cell-face velocity not only provides a more accur-
ate flux approximation (compared to the upwind and central schemes), but
also makes the homogeneous linear flux approximation consistent with Roe’s
approximation of the inviscid Burgers equation. The inclusion of the source
terms is shown to provide a more accurate flux approximation. For a uniformly
second-order accurate scheme, it is recommended that the source term is taken
to be piecewise constant.
Applying straightforward quadrature rules to the nonlinear integral repres-
entations does not result in a consistent approximation. The nonlinear ho-
mogeneous flux approximation breaks down in the inviscid limit and even for
diffusion dominated flows is less accurate than the linearized flux approxima-
tion. It is essential for the flux approximation to be consistent in the inviscid
limit for all cases. In Chapter 7, we formulate a new flux approximation scheme
using nonlinear local BVPs such that the numerical flux always reduces to the
Godunov flux in the inviscid limit.
75
Chapter 6
Complete-flux scheme for the

Navier-Stokes equations
In Chapter 4, it was shown that the discretization of the convective terms

in the incompressible Navier-Stokes equations using local BVPs, results in a
uniformly second-order accurate scheme that is free of spurious oscillations.
Further in Chapter 5, we presented the use of local BVPs for the approximation
of the fluxes in the semi-discrete viscous Burgers equation. Both approximation
methods are inspired by the complete-flux scheme for the convection-diffusion-
reaction equation described in [87], in which an integral representation for the
flux is derived using local BVPs for the entire equation, including the source
term.
We now extend the complete-flux scheme to the incompressible Navier-
Stokes equations. For the Navier-Stokes equations, we include the discrete
pressure gradient and the gradient of the cross-flux as source terms in the local
BVP, such that the flux approximation depends directly on the convection-
diffusion operator and the source terms. The nonlinear convective term causes
the local BVP to be nonlinear. To simplify the flux approximation, the local
BVP is linearized using cell-face velocity components. The cell-face velocity
components are in turn computed using local BVPs as presented in Chapter 4
and [48, 49]. The linearized local BVP is then integrated to obtain the flux
approximation, which can be expressed as the sum of a homogeneous part, cor-
responding to the convection-diffusion operator, and an inhomogeneous part,
corresponding to the source terms. To have a uniformly second-order accurate
approximation it is essential to include the source terms in the local BVP.
The chapter is organized as follows: in Section 6.1 we outline the underlying
finite-volume discretization. Section 6.2 gives the integral representations of the
fluxes as well as the closure of the flux scheme. In Section 6.3 we present the
numerical results for the flow in a lid-driven cavity and for the Taylor-Green
77
Complete-flux scheme for the Navier-Stokes equations
u,y
Ωvi,j+1/2 fi+1/2,j+1/2
yj+1
Ωui+1/2,j
vi,j+1/2
∆y u,x fi+1/2,j u,x
fi+1,j
fi,j
yj ×
pi,j ui+1/2,j
xi−1/2 xi+1/2 xi+3/2

u,y
∆x fi+1/2,j−1/2
Figure 6.1: (a) The staggered grid for the spatial discretization. (b) A control
volume Ωui+1/2,j for the spatial discretization of the u-momentum equation.
vortex problem. Finally, the conclusions are presented in Section 6.4.
6.1 Finite volume method

In this section we briefly outline the finite-volume discretization of the incom-
pressible Navier-Stokes equations, with the momentum equations formulated
as a balance of the fluxes. Consider the two-dimensional incompressible Navier-
Stokes equations
∇ · u = 0, (6.1a)

ut + ∇ · uu − ∇u = −∇p, (6.1b)
where u = u(x, y, t)ex +v(x, y, t)ey is the flow velocity, p(x, y, t) is the kinematic
pressure and = 1/Re, with Re being the Reynolds number of the flow. For
the spatial discretization we use a staggered grid configuration as shown in
Figure 6.1. We have different control volumes for the discretization of the u-
and v-momentum equations. We now express the momentum equations as a
balance of the fluxes, i.e.,
ut + ∇·f u = −∇ · (p ex ), f u := (u2 − ux )ex + (uv − uy )ey , (6.2a)

v v 2
vt + ∇·f = −∇ · (p ey ), f := (uv − vx )ex + (v − vy )ey . (6.2b)
Integrating equation (6.2a) over the control volume Ωui+1/2,j and applying
Gauss’ theorem we get
Z I I
ut dA + f u · n ds = − p ex · n ds,
Ωu
i+1/2,j
∂Ωu
i+1/2,j
∂Ωu
i+1/2,j
78
where n is the outward unit normal vector to the boundary ∂Ωui+1/2,j and
where the contour integral is oriented counter clockwise. This integral form of
the conservation law can be approximated over the control volume Ωui+1/2,j as
shown in Figure 6.1 using the mid-point rule as follows :
u,x u,x u,y u,y
∆y fi+1,j − fi,j +∆x fi+1/2,j+1/2 − fi+1/2,j−1/2 =
(6.3)
− ∆y pi+1,j − pi,j − ∆x ∆y (ut )i+1/2,j ,
u,x
where f u,x := u2 − ux , f u,y := uv − uy and fi,j ≈ f u,x (xi , yj ). Similarly,
equation (6.2b) can be discretized over the control volume Ωvi,j as :
v,x v,x v,y v,y
∆y fi+1/2,j+1/2 −fi−1/2,j+1/2 + ∆x fi,j+1 − fi,j =
(6.4)
− ∆x pi,j+1 − pi,j − ∆x ∆y (vt )i,j+1/2 ,
with f v,x := uv − vx and f v,y := v 2 − vy .
u,x
We begin with the approximation of the flux fi+1,j using the quasi-one-
dimensional formulation of equation (6.2a), i.e.,
(f u,x )x = s, s := −px − (f u,y )y − ut . (6.5)
Restricting the above equation to the interval x ∈ [xi+1/2 , xi+3/2 ] and y = yj ,
the boundary conditions read
u(xi+1/2 , yj ) = ui+1/2,j , u(xi+3/2 , yj ) = ui+3/2,j .
The term s acts as the source term for the flux f u,x , giving the forces generating
the flux. We have included the inertial term ut in the source term. However,
in this chapter we focus on the steady computation of the fluxes, i.e., ut = 0.
The components of the fluxes f u and f v are nonlinear, thereby making the
local BVPs nonlinear. The flux components are linearized using the interface
velocities which are computed at the interface of the control volume. For
example, the nonlinear term u2 in f u,x is linearized as U u, where U is the
approximation of the interface velocity. The details regarding the iterative
computation of the interface velocities using local BVPs are given in Chapter 4.
u,x
Thus, for the approximation of the flux fi+1,j we solve the linearized local BVP
(U u − ux )x = −px − (f u,y )y , x ∈ [xi+1/2 , xi+3/2 ], y = yj , (6.6a)

u(xi+1/2 , yj ) = ui+1/2,j , u(xi+3/2 , yj ) = ui+3/2,j . (6.6b)
In the next section we give the details regarding the solution of the above local
BVP.
6.2 Integral representation of the fluxes

The flux approximation scheme is based on the computation of the flux for
the scalar convection-diffusion-reaction equation as described in [87] and also
79
presented in Chapter 5. The model equation is given by

φt + (aφ − φx )x = s,
where the scalar flux is defined as f = a φ− φx , φ being the unknown variable.
We outline the computation of the scalar flux using a local BVP and then
extend the scheme to the Navier-Stokes equations. The computation of the
flux fi+1 at the cell edge xi+1 = 12 (xi+1/2 + xi+3/2 ) is based on the following
model BVP:
(a φ − φx )x = s, xi+1/2 < x < xi+3/2 , (6.7a)
φ(xi+1/2 ) = φi+1/2 , φ(xi+3/2 ) = φi+3/2 . (6.7b)
For the solution of the above local BVP we need the following variables:
Z x Z x
a
λ = , P = λ∆x, Λ(x) = λ(ξ)dξ, S(x) = s(ξ)dξ,
xi+1 xi+1
with ∆x = xi+3/2 − xi+1/2 . From Chapter 5, we get that the flux fi+1 is the
sum of a homogeneous (f h ) and an inhomogeneous (f inh ) part, i.e.,
h inh
fi+1 = fi+1 + fi+1 , (6.8a)
Z xi+3/2
h
fi+1 = e−Λi+1/2 φi+1/2 − e−Λi+3/2 φi+3/2 / e−Λ dx, (6.8b)
xi+1/2
Z xi+3/2 Z xi+3/2
inh
fi+1 =− e−Λ S dx / e−Λ dx. (6.8c)
xi+1/2 xi+1/2
For the incompressible Navier-Stokes equations, we first linearize the flux com-
ponent f u,x by defining f l,u,x = U u − ux , which can be computed using the
model BVP (6.7). In the following we restrict ourselves to the approximation of
the linearized flux f l,u,x . Now, the source term is given by s = −px − (f l,u,y )y .
To simplify the computation of the inhomogeneous part, we make the following
assumptions for the source term:
1. Pressure gradient: The pressure p is taken to be piecewise linear, con-
sequently the pressure gradient is piecewise constant, given by its central
approximation:
(
1
(δx p)i+1/2,j = ∆x (pi+1,j − pi,j ), xi+1/2 ≤ x ≤ xi+1 ,
px (x, yj ) ≈ 1
(δx p)i+3/2,j = ∆x (pi+2,j − pi+1,j ), xi+1 < x ≤ xi+3/2 .
2. Cross-flux gradient: the gradient of the cross-flux (f u,y )y is taken to be

piecewise constant, given by the central approximation:
(
u
u,y Ci+1/2,j , xi+1/2 ≤ x ≤ xi+1 ,
(f )y (x, yj ) ≈ u
Ci+3/2,j , xi+1 < x ≤ xi+3/2 ,
80
with
u 1 u,y u,y

Ci+1/2,j = Fi+1/2,j+1/2 − Fi+1/2,j−1/2 ,
∆y
F u,y being the numerical approximation of the linearized flux component
f l,u,x .
From the above we conclude that s is piecewise constant over the domain and
consequently S is piecewise linear. Moreover, we also have that U and are
constants on the domain (xi+1/2 , xi+3/2 ). Evaluating the expressions (6.8b)
u,x
and (6.8c), we get that the numerical flux Fi+1,j is given by
u,x u,x,h u,x,inh

Fi+1,j = Fi+1,j + Fi+1,j , (6.9a)
u,x,h
B(−P u )ui+1/2,j − B(P u )ui+3/2,j ,

Fi+1,j = (6.9b)
∆x
u,x,inh
= ∆x C(−P u )si+1/2,j − C(P u )si+3/2.j ,

Fi+1,j (6.9c)
where P u = U ∆ x/ and
z ez/2 − 1 − z/2
B(z) = , C(z) = .
ez − 1 z(ez − 1)
We further split the inhomogeneous part into terms depending on the gradi-
ent of the cross-flux term (F u,x,c ) and the pressure gradient (F u,x,p ), using
u
si+1/2,j = −(δx p)i+1/2,j − Ci+1/2,j in equation (6.9c).
u,x
Thus, we have computed the numerical flux Fi+1,j as the sum of a homo-
geneous and an inhomogeneous part using local BVPs. As shown in Section 5.1
the homogeneous flux component can be expressed as a weighted average of the
central flux (F cen ) and the upwind flux (F up ) as follows
F h = 2W1 (|P u |) F cen + 1 − 2W1 (|P u |) F up ,

where the weight function is given by W1 (z) = (ez −1−z)/z(ez −1). Again, for
diffusion-dominated flows (|P u | 1), the homogeneous scheme reduces to the
central scheme, whereas for convection dominated flows (|P u | 1), it reduces
to the upwind scheme. Analogously, the discrete source terms si+1/2,j and
si+3/2,j involved in the inhomogeneous part have equal contributions, when
|P u | 1. For higher Péclet numbers the upwind source term has a larger
contribution to the inhomogeneous flux part.
6.2.1 Closure of the scheme

So far we have derived an expression for the approximation of the flux com-
u,x
ponent Fi+1,j . For the semi-discrete momentum equation (6.3) we also need
to approximate the cross-flux f u,y . For the closure of the scheme we restrict
81
ourselves to the homogeneous flux part for the cross-flux component. Thus,
u,y
the flux Fi+1/2,j+1/2 is computed from the local BVP :
(V u − uy )y = 0, x = xi+1/2 , yj ≤ y ≤ yj+1 ,
u(xi+1/2 , yj ) = ui+1/2,j , u(xi+1/2 , yj+1 ) = ui+1/2,j+1 ,
where V is the estimate of the interface velocity at (xi+1/2 , yj+1/2 ). On solving
the above homogeneous local BVP we find that the flux is given by :
u,y
B(−P v )ui+1/2,j − B(P v )ui+1/2,j+1 , P v = V ∆y/.

Fi+1/2,j+1/2 =
∆y
Similarly, the fluxes and the cross-fluxes in equation (6.4) are also computed
v,y
from local BVPs. The flux fi,j+1 is computed from the inhomogeneous local
BVP :
(V v − vy )y = −py − (f v,x )x , x = xi , yj+1/2 ≤ y ≤ yj+3/2 ,
v(xi , yj+1/2 ) = vi,j+1/2 , v(xi , yj+3/2 ) = vi,j+3/2 ,
for which expressions analogous to (6.9) can be derived. Again, for the com-
v,x
putation of the cross-flux fi+1/2,j+1/2 we solve the homogeneous local BVP :
(U v − vx )x = 0, xi ≤ x ≤ xi+1 , y = yj+1/2 ,
v(xi , yj+1/2 ) = vi,j+1/2 , v(xi+1 , yj+1/2 ) = vi+1,j+1/2 .
In the following section we test these flux approximation schemes for the
lid-driven cavity and the Taylor-Green vortex problem.

6.3.1 Lid-driven cavity
For the lid-driven cavity, we use the results from Ghia-Ghia-Shin [29] as the
reference. Figure 6.2 and Figure 6.3 show the u-velocity profile along the
vertical center-line of the cavity. In Figure 6.2 we compare the homogeneous
flux scheme, the 1-D flux scheme (including the pressure-gradient as the source
term), and the 2-D flux scheme (including both cross-flux and the pressure-
gradient in the source term), computed on a coarse 20 × 20 grid, with the finer
grid (128 × 128) Ghia-Ghia-Shin solution for Re = 100. Since the pressure
gradient is practically zero across the domain, we do not see much difference
between the homogeneous and the 1-D flux scheme. However the inclusion of
the cross-flux term in the source term gives us a better approximation.
Next, we compare the flux scheme with the central scheme for Re = 400, (see
Figure 6.3). We compare the numerical solutions computed on a coarse 20 × 20
grid with the Ghia-Ghia-Shin results. Figure 6.3 also shows the comparison
between the flux approximation schemes and the central scheme. It can be
observed that the inhomogeneous flux schemes and the central scheme have
similar velocity profiles.
82
Figure 6.2: u-velocity profiles along the vertical centerline of the cavity for
Re = 100.
0.8
Reference
0.6 Central scheme
Homogeneous flux scheme
1D flux scheme
0.4
2D flux scheme
0.2
0
-0.4 0 0.4 0.8 1.2
Figure 6.3: u-velocity profiles along the vertical centerline of the cavity for
Re = 400.
83
Method Flux approx. Convective discretization
Central 5.8923 × 10−4

Upwind 1.1808 × 10−2
Homogeneous local BVP 8.7786 × 10−3 9.2407 × 10−3
Inhomogeneous 1-D local BVP 5.8923 × 10−4
2-D local BVP 5.8923 × 10−4
Table 6.1: Absolute errors (keu k∞ ) for the Taylor-Green vortex problem, for
Re = 100, on a 20 × 20 grid, at t = 20.
6.3.2 Taylor-Green vortex

We consider the Taylor-Green vortex problem for Re = 100 on the domain
[ 14 , 2 14 ] × [ 14 , 2 14 ] with periodic boundary conditions. For time integration, we
use the standard fourth-order, four-stage explicit Runge-Kutta method, with
a fixed time step ∆t = 10−3 . Figure 6.4 shows the convergence of the errors
keu k∞ = kunum − uexact k∞ over a family of uniform grids ranging from 5 × 5
to 320 × 320. The reference solution uexact is obtained by restricting the exact
solutions (4.25) of the Taylor-Green vortex problem to the computational grid.
It can be observed that Figure 6.4 is almost equal to Figure 4.8 with the cent-
ral, inhomogeneous 1-D and 2-D flux scheme exhibiting uniform second-order
convergence and the upwind scheme being uniformly first-order accurate. The
homogeneous flux scheme, as expected, is first-order accurate over coarse grids
and second-order accurate on finer grids. In fact, the errors for the 1-D flux
scheme and 2-D flux scheme are equal to the errors for the inhomogeneous 1-D
and 2-D local BVP schemes for the convective discretization. The absolute
error for the homogeneous flux scheme is a little smaller, compared to the ho-
mogeneous 1-D local BVP scheme for the convective discretization, also shown
in Table 6.1, without any noticeable changes in the computational cost. This
difference between the homogeneous flux scheme (HFS) and homogeneous 1-D
local BVP scheme for the convective discretization can be explained by the
lower numerical diffusion of the former as demonstrated in Chapter 5 for the
convection-diffusion equation (see Example 5.1).
6.4 Conclusion
In the preceding sections we presented methods for the approximation of the
fluxes derived from local BVPs. The computed flux is the sum of a homo-
84
10 0
Upwind
10 -2
keu k∞
Homogeneous flux
2
10 -4
Central
1-D flux
2-D flux
10 -6
10 -3 10 -2 10 -1 10 0
h
Figure 6.4: Convergence of the spatial errors for the Taylor-Green vortex prob-
lem, for Re = 100 and at t = 20, for a family of grids (5 × (2i , 2i ), i =
0, 1, 2, · · · , 6) using the standard explicit four-stage Runge-Kutta method with
∆t = 10−3 .
85
geneous and an inhomogeneous part. The homogeneous flux corresponding to

the convection-diffusion operator is a weighted mean of the upwind and central
scheme. The inhomogeneous part of the flux depends directly on the source
term in the BVP, i.e., the pressure-gradient and the cross-flux gradient. Com-
pared to the local BVP schemes for convective discretization, the homogeneous
flux scheme appears to be slightly more accurate than the homogeneous 1-D
local BVP approximation for the convective terms. The inclusion of the source
terms makes the flux approximation uniformly second-order convergent as ob-
served from the numerical results.
86
Chapter 7
Nonlinear flux approximation
In Chapter 5 we presented the approximation of the flux function for the viscous
Burgers equation using a nonlinear local BVP. Applying straightforward quad-
rature rules to the integral representation does not result in a consistent nu-
merical flux, as the approximation is shown to breakdown in the inviscid limit.
On the other hand, the numerical flux obtained using linearized local BVPs
is shown to reduce to Roe’s approximation for the inviscid Burgers equation
instead of the Godunov approximation, as a consequence of the linearization.
In this chapter we present a different approach for the approximation of the
flux using homogeneous nonlinear local two-point BVPs for the viscous Burgers
equation. A discussion on nonlinear local two-point BVPs can be found in [22],
where the authors show i) the solvability of some auxiliary local nonlinear two-
point BVPs, and ii) the convergence of the discrete scheme to a weak solution
of the continuous problem. We make use of the monotonicity of the solution
of the homogeneous local BVP to derive the numerical flux. Consequently, we
have to separately examine the cases i) uL ≥ uR , and ii) uL < uR , where uL
and uR are the left and the right boundary values, respectively. Subsequently,
we establish the existence and uniqueness of the numerical flux obtained from
the nonlinear local BVP. Since the numerical scheme preserves the nonlinearity
of the flux function, the nonlinear flux approximation scheme is shown to be
consistent with the Godunov approximation in the inviscid limit.
The chapter is organized as follows: in Section 7.1 we formulate the local
BVP for the flux approximation and show that the BVP has monotonic solu-
tions. Section 7.2 gives details of the derivation for the numerical fluxes. In
Section 7.3 we demonstrate the existence and uniqueness of the numerical flux
and show that the numerical flux reduces to the Godunov approximation in
the inviscid limit. In Section 7.4 we compare the nonlinear scheme with the
linearized homogeneous flux scheme (described in Chapter 5) as well as with
other standard methods. Section 7.5 gives the concluding remarks.
87
7.1 Flux from local two-point BVP

Consider the one-dimensional viscous Burgers equation
ut + f (u, ux )x = 0, f (u, ux ) := 12 u2 − νux , (7.1)
defined on Ω(⊂ R)×(0, T ), where ν(≥ 0) is the diffusion coefficient. The spatial
discretization of the Burgers equation using a finite-volume method requires
the approximation of the flux function f (u, ux ) at each interface between two
control volumes. The semi-discrete formulation of equation (7.1) is given by
h u̇i + Fi+1/2 − Fi−1/2 = 0,
where Fi+1/2 ≈ f (u, ux )|x=xi+1/2 , see Figure 7.1. The derivation of the flux
Fi+1/2 is based on the following model BVP, in which we ignore the time
derivative
fx = 12 u2 − νux x = 0, x ∈ (xi , xi+1 ),

(7.2a)
u(xi ) = ui = uL , u(xi+1 ) = ui+1 = uR . (7.2b)
The solution of the nonlinear BVP (7.2) provides us the numerical flux function
F(uL , uR , ν/h), which is constant on the interval (xi , xi+1 ). Thus:
Fi+1/2 = F(uL , uR , ν/h).
Using the normalized coordinate σ, σ ∈ [0, 1] and the parameter , defined by

x − xj ν
σ := , := ,
h h
the BVP (7.2) can be expressed as
1 2

2u − uσ σ
= 0, σ ∈ (0, 1), (7.3a)
u(0) = uL , u(1) = uR . (7.3b)
Further, it can be shown that the above BVP has a monotonic solution.
Fi−1/2 Fi+1/2
ui−1 ui ui+1 ui+2

xi−1 xi xi+1/2 xi+1
h
Figure 7.1: Spatial discretization for the one-dimensional Burgers equation
88
Lemma 7.1. The nonlinear local boundary value problem (7.3) has a strictly
monotonic solution. Further, it can be shown that the derivative uσ satisfies
U U
(uR − uL ) exp − (1 + σ) ≥ uσ (σ) ≥ (uR − uL ) exp (1 + σ) , (7.4)

if uL ≥ uR with U := max{|u(η)|, η ∈ [0, 1]}, otherwise for uL < uR ,
U U
(uR − uL ) exp − (1 + σ) ≤ uσ (σ) ≤ (uR − uL ) exp (1 + σ) . (7.5)

Proof. Any solution u of the problem (7.3) can be represented as
Λ(σ) λ(σ)
u(σ) = uL + (uR − uL ) , uσ (σ) = (uR − uL ) ,
Λ(1) Λ(1)
for σ ∈ [0, 1] (see Lemma 3.1, [22]), where the functions λ, Λ : [0, 1] → R are
given by
1 Z σ Z σ
λ(σ) := exp u(η)dη and Λ(σ) := λ(ξ)dξ.
0 0
Clearly, for uL > uR , uσ (σ) < 0 causing u(σ) to be a monotonically decreas-

ing function. Similarly, if uL < uR , then uσ (σ) > 0 and u is monotonically
increasing. Further, note that exp(−U σ/) ≤ λ(σ) ≤ exp(U σ/). Similarly,
for the integral Λ(1), we have exp(−U/) ≤ Λ(1) ≤ exp(U/). From the above
inequalities for λ(σ) and Λ(1), we get as the bounds for the derivative uσ (σ):
(7.4) and (7.5).
7.2 Numerical flux function

We now derive expressions for the numerical flux function using the BVP (7.3).
As a consequence of Lemma 7.1 we consider the cases: uL ≥ uR and uL < uR .
7.2.1 The case uL ≥ uR

The solution of the BVP (7.3) in this case results in a (strictly) decreasing
function, i.e., uσ ≤ 0. Using the left boundary condition u(0) = uL , we get
that the numerical flux at the interface, Fi+1/2 , is given by
1 2
Fi+1/2 = f (0) = u − uσ (0).
2 L
Alternatively, the flux can be determined using the right boundary condition
1 2
Fi+1/2 = u − uσ (1).
2 R
89
Since uσ < 0, we conclude that Fi+1/2 > 0, therefore there exists a c ∈ R, such
that
1 1 1
Fi+1/2 = u2 − uσ = u2L − uσ (0) = 12 u2R − uσ (1) = c2 ,
2 2 2
with |c| ≥ max(|uL |, |uR |). The above relation gives us the first-order differen-
tial equation with unknown parameter c
1 2
uσ = (u − c2 ), σ ∈ (0, 1),
2
which needs to satisfy both u(0) = uL and u(1) = uR . Integrating the differen-
tial equation and connecting the left and the right boundary condition results
in the following nonlinear equation for the unknown c with parameters uL , uR
and (u + c)(u − c) c
L R
H + (c) := log − = 0. (7.6)

(uL − c)(uR + c)
Thus, Fi+1/2 is given by the non-trivial roots of the function H + (c), which
is an odd function. We restrict ourselves to c > 0. Note that the nonlinear
equation (7.6) can also be expressed as
e−c/2 (uL + c)(uR − c) − ec/2 (uL − c)(uR + c) = 0.

(7.7)
Let s = (uL + uR )/2, then for s ≥ 0, we get that uL ≥ |uR | and the non-trivial
solution of equation (7.7) satisfies c ≥ uL ≥ |uR |. In the inviscid limit → 0,
for s ≥ 0 equation (7.7) reduces to
ec/2 (c − uL )(c + uR ) = 0 ⇒ c = uL .
Similarly, for s < 0, we have uR < 0, implying c ≥ −uR ≥ |uL | and the limit
case solution is then given by c = −uR (> 0). Thus, the numerical flux in the
inviscid limit is given by

 1 u2L , if s ≥ 0,
2
Fi+1/2 =
 1 u2 , if s < 0,
2 R
which is actually the Godunov flux for the inviscid Burgers equation [31, 93].
Moreover, if uL = uR = u, then uσ = 0 and the numerical flux is given by
Fi+1/2 = F(u, u) = 21 u2 = f (u), for constant u. Hence the numerical flux
function F is consistent with the continuous flux function f .
In order to compute the numerical flux Fi+1/2 , we need to compute the
non-trivial zeros of the function H + (c), which are given by equation (7.6) or
equivalently, by equation (7.7). The nonlinear equation can be solved using
Newton iteration, given a good initial guess for the flux. Algorithm 7.1 shows
the steps for choosing a good initial guess for the Newton solver using the
bounds of the derivative uσ as shown in Lemma 7.1.
Next, we consider the case uL < ur .
90
Algorithm 7.1: Choosing an initial guess for the numerical flux:

Fi+1/2 = c2 /2
Given: uL , uR ,
Set s = (uL + uR )/2

 uL if s ≥ 0,
Set U :=
|u | if s < 0.
R
Using (7.4) as a bound for uσ (0)


 12 u2L + (uL − uR ) exp − U if s ≥ 0,
Set f0 :=
 1 u2 + (u − u ) exp − U if s < 0.
2 R L R
√
Initial guess: c0 = | 2f0 |
7.2.2 The case uL < uR

From Lemma 7.1 we conclude that uσ > 0 for uL < uR . Thus Fi+1/2 =
u2 /2 − uσ > 0 if uσ < u2 /2 and Fi+1/2 < 0 if uσ > u2 /2. Therefore, we split
the derivation of the numerical flux into three cases, depending on the sign of
the flux.
Case 0: Zero-flux condition

Before we examine the cases where the flux is positive or negative, we derive
the condition for zero-flux across the interface. We have established that the
numerical flux Fi+1/2 is given by
1 2 1
Fi+1/2 = c = u2 − uσ .
2 2
If the net flux through the interface is zero, then we have
1 2 du
u = .
2 dσ
Integrating the above differential equation and applying the left boundary con-
dition gives
1 1 σ
− = .
uL u(σ) 2
Ensuring that the above relation also satisfies the right boundary condition,
u(1) = uR , results in the zero-flux condition :
1 1 1
− = (uL , uR , 6= 0). (7.8)
uL uR 2
Note that relation (7.8) can hold for both uL < uR and uL ≥ uR .
91
Case 1: Positive flux
If the flux is positive, then the numerical flux is evaluated as for the case
uL ≥ uR and is given by roots of the function H + (c), defined in equation (7.6),
with c ∈ (0, min(|uL |, |uR |)).
Case 2: Negative flux
If the flux is negative, then there exists a c ∈ R, such that
1 2 1
Fi+1/2 = u − uσ = − c2 .
2 2
This relation gives rise to the first-order differential equation
1 2
uσ = (u + c2 ), σ ∈ (0, 1), (7.9)
2
with the boundary conditions (7.3b). Integrating the first-order differential

equation and connecting the left and right boundary condition gives us another
nonlinear equation for c, i.e.,
u u c
R L
H − (c) := arctan − arctan − = 0. (7.10)
c c 2
As before, the numerical value of Fi+1/2 = −c2 /2 is given by the non-trivial

roots of the function H − (c). Since H − (c) is an odd function, we restrict
ourselves to c > 0.
So far, we have considered different cases for determining the numerical flux
depending on uL , uR and the sign of the flux. For the practical application of
the flux scheme it is essential that the flux is always unique, whenever it exists.
In the following section, we show that for both uL ≥ uR and uL < uR , the
numerical flux is unique.
7.3 Existence of numerical flux

In this section we show the existence and uniqueness of the numerical flux
given by the roots of H + (c) and H − (c). We begin with the case uL < uR ,
where the flux can be either positive, negative or zero and present a criterion
for identifying the sign of the flux.
Lemma 7.2. For uL , uR , 6= 0, and uL < uR , the sign of the flux can be
determined as follows
92
If uL uR > 0, then

− 1 c2 if 1
uL − 1
uR > 1
2 ,
 2



1 2 1 1 1
Fi+1/2 = 2c if uL − uR < 2 ,
(7.11)


1 1 1

 0 if uL − uR = 2 .
If uL uR < 0, then
1
Fi+1/2 = − c2 .
2
The numerical flux can then be computed using the roots of the function,
H + (c) if Fi+1/2 > 0 and H − (c) if Fi−1/2 < 0.
Proof. We have already established the condition for zero-flux, i.e., relation
(7.8). In the following we focus on the non-trivial roots of the functions H + (c)
and H − (c). Let α− (c) := arctan(uR /c)−arctan(uL /c) and β− (c) := c/2, such
that H − (c) := α− (c) − β− (c). Using arctan(1/z) = − arctan(z) + πsgn(z)/2,
α− (c) can be expressed as
c c π
α− (c) = arctan − arctan + sgn(uR ) − sgn(uL ) ,
uL uR 2
with
0 uL uR (uR − uL )(uL uR − c2 ) d
α− (c) = − 2 = , (.)0 := .
u2L+c2 uR + c2 (u2L + c2 )(u2R + c2 ) dc
Using the fact that α− (c) is an odd function we restrict ourselves to the case
c > 0.
We split the proof into the following cases :
• 0 < uL < uR .
√
Observe that α− (c) is an increasing function for c ∈ (0, uL uR ] with a
√
maximum at c = uL uR (< uR ). Clearly H − (c) has a non-trivial root
whenever α− (c) = β− (c), i.e., the two functions intersect for c > 0, which
0 0
is possible only if α− (0) > β− (0), or,
0 1 1 0 1
α− (0) = − > β− (0) = .
uL uR 2
Thus, if the above condition holds then H − (c) has a non-trivial solution,
√
which satisfies uL uR < c < uR . Next, we investigate the condition
under which H + (c) has a non-trivial root. Let
(uR − uL )c
z(c) := ,
uL uR − c2
93
such that z ∈ [0, 1] with z(0) = 0, z(uL ) = 1. Clearly, for c ∈ (0, uL ) we

have z(c) > 0. Using z(c) we can rewrite equation (7.6) as
1 + z(c) c c
H + (c) = log − = 2 Artanh(z(c)) − , c ∈ (0, uL ).
1 − z(c)
Further, let α+ (c) := 2 Artanh(z(c)) and β+ (c) := c/, such that H + (c) =
α+ (c) − β+ (c) with
0 1 uL uR + c2 0 1
α+ (c) = 2(uR − uL ) , β+ (c) = .
1 − z 2 (c) (uL uR − c2 )2
For c ∈ (0, uL ), z(c) ∈ (0, 1) implying that α+ (c) is an increasing function,

0 0
thus H + (c) has a non-trivial root only if α+ (0) < β+ (0). The condition
0 0
α+ (0) < β1 (0) translates to
1 1 1
− < .
uL uR 2
• uL < uR < 0.
Again, we restrict ourselves to c > 0. For uL < uR < 0, we have c ∈
(0, |uR |), with |uR | < |uL |, which gives us that c2 < uL uR , resulting in
0
α− (c) > 0. Moreover, it can be shown that the function α− (c) is an
√ √
increasing function on (0, uL uR ), with a local maximum at c = uL uR
0 √ 0
and α− (c) < 0, for c > uL uR . Further, β− (c) = 1/2 > 0, with
β− (c) → ∞, in the limit c → ∞. Thus, for H − (c) to have a non-trivial
0 0
root, it is necessary that α− (0) > β− (0), which is equivalent to
1 1 1
− > .
uL uR 2
Similarly, for H + (c) to have a non-trivial root in (0, |uR |), requires α+ (c) =
0
β+ (c). Since uL uR > 0, we get that α+ (c) > 0, for c ∈ (0, |uR |), causing
α+ (c) to be monotonically increasing. With α+ (0) = β+ (0) = 0 and
α+ (c) → ∞ in the limit c → ∞, we get that H + (c) has a non-trivial root
0 0
in the interval (0, |uR |), if α+ (0) < β+ (0), which is is equivalent to
1 1 1
− < .
uL uR 2
• uL < 0 < uR .
Let umin := min(|uL |, uR ), giving us c ∈ (0, umin ). For c ∈ (0, umin ),
z(c) < 0 with z(umin ) = −1. Therefore,
1 + z(c)
0
α+ (c) = log < 0, with α+ (c) < 0.
1 − z(c)
94
0
However, β+ (c) > 0, with β+ (c) > 0. Thus, we get that H + (c) has no
non-trivial solution, whenever uL uR < 0. Further, if uL < 0 < uR , then
1 1 1
− < 0 6= ,
uL uR 2
implying that the flux can not be zero. Thus for uL < 0 < uR , the flux
can be neither zero nor positive, which leaves us with only one possibility,
i.e., the flux being negative and given by the non-trivial roots of H − (c).
For the function H − (c), we have that H − (0) = π, with
d − (uR − uL )(uL uR − c2 ) 1
H (c) = 2 2 2 2
− < 0.
dc (uL + c )(uR + c ) 2
Moreover, for c 1, β− (c) > α− (c) and H − (c) < 0. Thus, H − (c) always
has a non-trivial root for sufficiently large c. Hence, for uL < 0 < uR ,
H − (c) always has a non-trivial root.
Lemma 7.2 presents a criterion for identifying the sign of the numerical flux,
which in turn tells us whether the flux is given by the non-trivial roots of H + (c)
or H − (c). To demonstrate this, we consider the case uL = 0.25, uR = 1 for
= 10−1 , 1 and 10. For the given values of uL , uR and , 1/uL − 1/uR − 1/2 ∈
{−2, 2.5, 2.95}. From Lemma 7.2 we get that H + (c) will have a non-trivial
root only for = 10−1 , which can be observed in Figure 7.2. Moreover, for
= 10−1 , it can be observed that the non-trivial root of H + (c) is smaller
√
than uL uR = 0.5. For both = 1 and = 10, 1/uL − 1/uR > 1/2 and
consequently H − (c) has a non-trivial root, as shown in Figure 7.3. We have
also shown that for uL uR < 0, the flux is always negative and thus is given
by the non-trivial roots of H − (c). This is depicted in Figure 7.4, where we
consider the cases: uL = −0.25, uR = 1 and = 10−2 , 1 and 10. For = 10−2 ,
α− (c) = β− (c) for c ≈ 0.36 and the flux c2 /2 ≈ 0.06.
Lastly, to establish the existence of the numerical flux, we need to show
that the function H + (c) always has a non-trivial root, whenever uL > uR .
Lemma 7.3. For uL > uR , the function H + (c) always has a non-trivial root.
Proof. For uL > uR , we have shown that the flux is positive and is given by
the non-trivial roots of the function H + (c) = α+ (c) − β+ (c) = 0, where
(u + c)(u − c) c
L R
α+ (c) = log and β+ (c) = .

(uL − c)(uR + c)
For c ≥ cmin := max{|uL |, |uR |}, we get that
(uL + c)(uR − c)
P (c) := > 0,
(uL − c)(uR + c)
95
3 α+(c)
β+(c) (ǫ = 10−1 )
β+(c) (ǫ = 1)
β+(c) (ǫ = 10)
2
0
0 0.05 0.1 0.15 0.2 0.25
c
Figure 7.2: Function α+ (c) for uL = 0.25, uR = 1 and β+ (c) for = 10−2 , 1
and 10.
2
α−(c)
β−(c) (ǫ = 10−1 )
β−(c) (ǫ = 1)
β−(c) (ǫ = 10)
0
0 1 2 3 4 5
c
Figure 7.3: Function α− (c) for uL = 0.25, uR = 1 and β− (c) for = 10−2 , 1
and 10.
96
2
α−(c)
β−(c) (ǫ = 10−1 )
β−(c) (ǫ = 1)
β−(c) (ǫ = 10)
0
0 3 6
c
Figure 7.4: Function α− (c) and β− (c) for uL = −0.25, uR = 1 and = 10−2 , 1
and 10.
with
(c2 + uL uR )
P 0 (c) = −2(uL − uR ) < 0.
(uL − c)2 (uR + c)2
As c → cmin , we have P (c) → ∞, implying α+ (c) → ∞, and for c → ∞,
P (c) → 1, giving α+ (c) = 0, with
0 P 0 (c)
α+ (c) = < 0.
P (c)
Thus, over the domain (cmin , ∞), α+ (c) is a monotonically decreasing function
which asymptotically goes to 0 as c → ∞. Further, the function β + (c) = c/,
is a monotonically increasing function with β+ (cmin ) = cmin / and β+ (c) → ∞,
in the limit c → ∞. Therefore, H + (c) always has a unique non-trivial root, for
uL > uR .
7.3.1 The limit flux

For the case uL ≥ uR we have shown that in the limit ν → 0, the numerical
flux is consistent with Godunov’s approximation. We now determine the limit
flux for the case uL < uR .
In the limit → 0, the flux is given by Fi+1/2 = u2 /2. To determine u, it
is essential to examine the behavior of the solution u(σ) of the nonlinear local
BVP (7.3). Instead of computing the non-trivial roots of the function H + (c)
97
and H − (c), we solve the local BVP(7.3) numerically using the central scheme
over a very fine grid.
Figure 7.5 shows the numerical solutions of the nonlinear local BVP (7.3)
with the boundary conditions given by uL = 0.5 and uR = 1, for = 10−1 , 10−2
and 10−3 . Observe that for smaller values of , the solution has a boundary
layer character, i.e., u(σ) = uL all throughout the domain except in a very
small region close to σ = 1, where the solution adapts to the right boundary
condition. For the considered boundary conditions uL = 0.5 and uR = 1.0, we
1
1 ǫ = 10−1 ǫ = 10−3
ǫ = 10−2 ǫ = 10−4
ǫ = 10−3
u(σ)
u(σ)
0.5
0.5
0
0 0.5 1 0 0.5 1
σ σ
(a) (b)
Figure 7.5: Numerical solutions of the nonlinear local BVP (7.3) using the
central scheme on a very fine grid (∆σ = 10−4 ) for the boundary conditions:
(a) uL = 0.5 and uR = 1 for = 10−1 , 10−2 , 10−3 , (b) uL = 0 and uR = 1 for
= 10−3 , 10−4 .
have that H − (c) has non-trivial solutions provided > 0.5, thus in the limit
→ 0, one should analyze the non-trivial roots of the function H + (c)
u u + c(u − u ) − c2 c
L R R L
H + (c) = log −
uL uR − c(uR − uL ) − c2
1 + z(c)
= log exp(−c/) ,
1 − z(c)
where
(uR − uL )c
z(c) := .
uL uR − c2
Clearly, the zeros of H + (c) satisfy
exp(−c/2)(1 + z(c)) − exp(c/2)(1 − z(c)) = 0. (7.12)
98
For 0 < uL < uR and Fi+1/2 > 0,
1 2 1 1
Fi+1/2 = c = u2L − uσ (0) = u2R − uσ (1),
2 2 2
giving us that the non-trivial root of H + (c) is located in c ∈ (0, uL ]. Now as

→ 0, exp(−c/2) → 0 and equation (7.12) reduces to
exp(c/2)(1 − z(c)) = 0, (7.13)
which has a solution for z(c) = 1, i.e., c = uL and Fi+1/2 = u2L /2, which is in
agreement with the numerical solution of the BVP (7.3) shown in Figure 7.5.
Moreover, using the numerical solution of the nonlinear BVP (7.3), it can be
shown that for 0 = uL < uR , the solution is of boundary layer type with
u(σ) = uL = 0 as → 0, as shown in Figure 7.5(b).
Next, we consider the case uL < uR < 0, in which we have c ∈ (0, |uR |],
with z(c) ∈ (0, 1] and z(|uR |) = 1. Again in the limit → 0, the reduced
equation (7.13) has a solution for z(c) = 1, giving c = −uR and Fi+1/2 = u2R /2.
Figure 7.6 shows numerical solutions of the nonlinear local BVP (7.3) with
uL = −1.5, uR = −0.5 and = 10−1 , 10−2 and 10−3 . It can be seen that for
smaller values of , the solution of the BVP has a boundary layer character,
with u(σ) ≈ uR all through the domain, except in very small region close to
the left boundary, which is in agreement with the limit solution of the reduced
equation (7.13). As in the previous case, the numerical solution of the nonlinear
BVP (7.3) with uL < uR = 0, also exhibits a boundary layer character (for
small ) with u(σ) ≈ uR = 0, except in a very small region close to the left
boundary.
In Lemma 7.2, we showed that for uL < 0 < uR , the flux can not be positive
as the function H + (c) does not have a non-trivial root. On the other hand, if
H − (c) has a root then, α− (c) = β− (c). In the limit → 0, β− (c) = c/2 → ∞.
With α− (c) being a monotonically decreasing function and α− (0) = π, the
only root of H − (c), is c = 0, resulting in Fi+1/2 = 0. This is better explained
using the numerical solution of the nonlinear local BVP (7.3) computed using
the central scheme on a very fine grid with ∆σ = 10−4 . Figure 7.7 shows
numerical solutions of the nonlinear BVP for = 10−2 and 10−3 . Unlike, in
Figure 7.5 and Figure 7.6, for uL < 0 < uR , the solution of the local BVP
(7.3) possesses two boundary layers with u(σ) = 0 for the rest of the domain.
For very small , u(σ) = 0, yielding a zero-flux passing through the interface
xi+1/2 .
The discussion for the limit case can be summarized as:
99
-0.5
-0.75
u(σ)
-1
-1.25 ǫ = 10−1
ǫ = 10−2
ǫ = 10−3
-1.5
0 0.5 1
σ
central scheme on a very fine grid (∆σ = 10−4 ) with uL = −1.5 and uR = −0.5.
0
u(σ)
-1
ǫ = 10−3
ǫ = 10−4
-2
0 0.5 1
σ
central scheme on a very grid (∆σ = 10−4 ) with uL = −2 and uR = 1.
100
For uL < uR , the flux in the limit → 0, is given by

1 2
Fi+1/2 = c ,
2
where c is given by:

uL
 if 0 ≤ uL < uR ,
c = uR if uL < uR ≤ 0,

0 if uL < 0 < uR .


We compare the proposed nonlinear local BVP scheme with the upwind scheme
and the linearized homogeneous flux scheme described in Chapter 5. Recall that
lin,h
for the linearized homogeneous flux scheme, the numerical flux Fi+1/2 is given
by
lin,h
Fi+1/2 = B(−P )uL − B(P )uR ,
where B(z) := z/(ez − 1) is the Bernoulli function and P := ui+1/2 /2 the
mesh Péclet number. For the interface velocity ui+1/2 , we take ui+1/2 = (uL +
uR )/2. We use the analytical solution to the viscous Burgers equation defined
on (0, 1) × (0, T ), T ∈ (0, 1]
1
uref (x, t) = 1 + 1
2 tanh x − 0.1 − 12 t , (7.14)
4ν
as the reference solution to compare the schemes. We use the standard explicit
fourth-order Runge-Kutta scheme for time integration with ∆t = 10−3 . For this
test case, we have 0 < uR < uL throughout the computational domain, and the
numerical flux is given by the roots of H + (c). The Newton solver converges
in 2 − 8 iteration steps (depending on the tolerance, ranging from 10−3 to
10−8 ), for a good initial guess. A fairly accurate initial guess can be derived
using the bounds on the derivative uσ , that can be obtained using Lemma 7.1.
Figure 7.8 shows the convergence of the absolute error keu k1 := ku − uref k1 for
ν = 10−3 over a family of uniform grids. Grid refinement (for fixed ν) causes
to increase (for the test case, = 2i × 10−2 , i = 1, 2, · · · , 6). Moreover, it
is observed that the root-finder converges faster for higher values of (≈ 1)
than for smaller values of . On coarse grids all three schemes exhibit first-
order accuracy, with the local BVP schemes being slightly more accurate than
the upwind scheme. However, on grid refinement (increasing ) the nonlinear
BVP scheme is found to be more accurate compared to the upwind and the
linearized local BVP scheme (Figure 7.8). For ν = 10−2 , i.e., for a smoother
101
||eu ||1 10 -2 1
2
10 -3
nonlinear local BVP

10 -4
linear local BVP
upwind
10 -3 10 -2 10 -1
h
Figure 7.8: Convergence of the 1-norm of the absolute error for ν = 10−3 for the
proposed nonlinear local BVP scheme, homogeneous linear local BVP scheme
and the upwind scheme, for a family of grids (h = 10−1 × 2−i ; i = 1, 2, · · · , 6).
velocity profile, the nonlinear local BVP scheme exhibits uniform second-order
convergence, as shown in Figure 7.9.
7.5 Conclusion
In this chapter, we have presented a flux approximation scheme for the viscous
Burgers equation, in which the numerical flux function is given by the solution
of a local nonlinear two-point BVP, resulting in a locally exact approximation
which corresponds with the nonlinearity of the flux function. The flux is com-
puted using a nonlinear local BVP, which has monotonic solutions and exhibits
a boundary layer character for smaller values of (= ν/h). Thus depending on
the boundary conditions uL , uR , the local BVP has either a monotonically
increasing or monotonically decreasing solution. For uL ≥ uR , the solution of
the local BVP is monotonically decreasing, with the flux being positive and
given by the roots of the function H + (c). For uL < uR , the solution of the
local BVP is monotonically increasing, thus uσ > 0. Hence, the flux can be
positive, negative or zero. We have presented a criterion for identifying the
sign of the flux depending on the boundary conditions uL , uR and . For both
cases uL ≥ uR and uL < uR , the numerical flux is shown to be consistent with
the Godunov method in the inviscid limit.
102
10 -2
2
||eu ||1
10 -4
nonlinear local BVP
linear local BVP
upwind
10 -6
10 -3 10 -2 10 -1
h
Figure 7.9: Convergence of the 1-norm of the absolute error for ν = 10−2 for the
proposed nonlinear local BVP scheme, homogeneous linear local BVP scheme
and the upwind scheme, for a family of grids (h = 10−1 × 2−i ; i = 1, 2, · · · , 6).
103
Chapter 8
Conclusion
8.1 Concluding remarks

In the previous chapters, we presented spatial discretization techniques for flow
problems involving convection and diffusion, using local BVPs. The spatial dis-
cretizations are designed such that the resulting scheme has a 3-point stencil
(in each spatial direction), is uniformly second-order accurate and does not in-
troduce any non-physical oscillations. Standard schemes like the second-order
accurate central scheme and the first-order upwind scheme result in a 3-point
discretization. The central scheme suffers from spurious oscillations whenever
the absolute value of the grid Péclet number is higher than 2 and hence can
be computationally quite expensive due to the grid size restriction. Unlike the
central scheme, the upwind scheme does not introduce any non-physical oscil-
lations, but it is only first-order accurate. Hybrid/exponential schemes which
are equivalent to the homogeneous local BVP approximation all result in a
positive discretization and are first-order accurate on coarse grids but second-
order accurate on finer grids. The use of local BVPs allows us to include the
effects of source terms, which is shown to reduce the numerical diffusion of the
homogeneous approximation making the scheme uniformly second-order accur-
ate. In the following we give the conclusions and recommendations for using
local BVPs for the spatial discretization.
Convective flux approximation. We demonstrated the use of local BVPs

for the discretization of the convective terms involved in the constant coefficient
convection-diffusion equation and the incompressible Navier-Stokes equations.
For the convection-diffusion equation, we showed that indeed the inclusion of
the source term, i.e., the discrete time derivative of the unknown, removes
numerical diffusion from the homogeneous part of the approximation. Thus,
105
Conclusion
the inhomogeneous local BVP approximation for the convection-diffusion equa-

tion results in a uniformly second-order accurate scheme, which is free of any
spurious oscillations, even on coarse grids.
For the incompressible Navier-Stokes equations, the convective term is non-
linear, which makes the local BVP nonlinear. In order to simplify the problem
and to derive explicit expressions for the approximation of the cell face velocity
components, the local BVP is linearized and then solved iteratively. Again, the
approximation is expressed as the sum of a homogeneous part, corresponding
to the local convection-diffusion balance, and an inhomogeneous part, corres-
ponding to the source terms included in the local BVP. The homogeneous local
BVP method, like the hybrid scheme, provides an exponential reconstruction
of the velocity component within the control volume and can be expressed as
a weighted sum of the central and the upwind scheme. For the inhomogeneous
part, we use the pressure gradient and the gradient of the transverse flux as
source terms. The numerical results show that the inhomogeneous approxim-
ation improves the accuracy of the homogeneous approximation. In the 1-D
inhomogeneous local BVP approximation, the pressure gradient is assumed to
be piecewise constant and in the 2-D inhomogeneous local BVP approxima-
tion, the gradient of the cross-flux is assumed to be constant along with the
piecewise constant pressure gradient. Both 1-D and 2-D inhomogeneous local
BVP schemes exhibit uniform second-order convergence and do not exhibit
any spurious oscillations, not even on coarse grids. Also, it is observed that
the inclusion of the source term reduces the dissipation of the kinetic energy
due to numerical diffusion (introduced by the homogeneous part). Since the
inhomogeneous local BVP schemes do not introduce any or very little numer-
ical diffusion, we expect the schemes to work well for simulation of wind farm
dynamics, particularly on coarse grids. Due to its higher accuracy at margin-
ally higher cost, the 2-D inhomogeneous local BVP is recommended.
Flux approximation using local BVPs. We extended the use of local

BVPs for the approximation of the net flux, i.e., the sum of the convective and
diffusive flux. In Chapter 5 we extended the complete-flux scheme, introduced
in [87], to the viscous Burgers equation. The nonlinearity of the flux function
for the viscous Burgers equation causes the model BVP for the flux approx-
imation to be nonlinear as well. The integral representations thus obtained
are implicit, i.e., depend on the solution of the BVP. To simplify the problem
and to derive explicit expressions for the numerical flux, the local BVP is lin-
earized using interface velocities. In order to ensure that the numerical flux
approximation is consistent in the inviscid limit, it is essential that the inter-
face velocity used for linearization is computed using the upwind scheme or the
1-D homogeneous local BVP scheme. The linearized local BVP scheme is also
extended for the approximation of the numerical fluxes in the incompressible
Navier-Stokes equations, with the piecewise constant discrete pressure gradi-
ent and the piecewise constant discrete gradient of the transverse flux as source
106
Conclusion
terms. As expected, the inclusion of the source terms provides a more accurate
approximation (uniformly second-order accurate) than the homogeneous local
BVP scheme, also here at only marginally higher cost.
Nonlinear flux approximation for the Burgers equation. The solution of

the nonlinear local BVP exhibits a thin boundary layer character in the inviscid
limit. Using a numerical solution computed on a very fine grid, it can be shown
that the homogeneous local BVP (the steady viscous Burgers equation) always
has a boundary layer in the inviscid limit. Further, if the boundary conditions
(uL and uR , see Chapter 7) are of opposite sign with uL < 0 < uR , then in the
inviscid limit the numerical solution of the homogeneous local BVP possesses
two boundary layers, with the net flux being zero. The linearized local BVP
approximation, which provides an exponential reconstruction of the solution of
the nonlinear local BVP, only exhibits one of the boundary layers. To derive a
numerical flux approximation, which is always consistent, it is essential to take
into account the nonlinearity of the flux function. Moreover, we showed that
applying standard quadrature rules to the nonlinear integral representations
of the flux does not result in a consistent approximation, as the resulting net
numerical flux reduces to zero.
We studied a new approach to compute the flux from the nonlinear local
BVP. Instead of determining the integral representation, the flux is computed
by the direct solution of the nonlinear local BVP. We showed that for this new
nonlinear local BVP scheme, in the inviscid limit the numerical flux reduces
to the Godunov flux for the inviscid Burgers equation. Further, we established
the uniqueness of the numerical flux and showed that in all cases (including
uL < 0 < uR ) the numerical flux reduces to the correct limit flux. The scheme
requires a nonlinear solver to compute the root of the derived equations, for
which a good initial guess is available (shown in Chapter 7).
8.2 Recommendations for future work

As shown in Chapter 7, the nonlinear local BVP scheme provides a numerical
flux approximation that is consistent in the inviscid limit. However, the scheme
is not uniformly second-order accurate; over coarse grids it displays first-order
accuracy. In order to ensure that the scheme is uniformly second-order accur-
ate, it is essential that the source term is included in the local BVP. One way
to include the source term in the local BVP is to show the equivalence of the
nonlinear flux approximation (presented in Chapter 7) with the integral rep-
resentation of the homogeneous part of the flux (presented in Chapter 5). This
can be established if we show that the homogeneous nonlinear local BVP has
a unique solution (as shown in [22] for an auxiliary problem). Then the non-
linear flux approximation can be combined with the integral formulation of the
inhomogeneous flux, such that the resulting scheme is uniformly second-order
107
Conclusion
accurate. Further, the nonlinear flux approximation may be combined with a

WENO reconstruction to get a higher-order flux approximation scheme. The
nonlinear local BVP scheme should be extended to more complex problems as
the incompressible Navier-Stokes equations.
In Chapter 7, we mainly focused on the derivation of the numerical ap-
proximation of the flux function f = u2 /2 − νux , with a nonlinear convective
part and linear diffusion. However, for some other problems which have a
somewhat different convection term, say g(u), the similar derivation of the nu-
merical flux as shown in Chapter 7 can be carried out provided the convection
term g(u) is convex. Further, for flow problems involving nonlinear diffusion,
i.e., the diffusion coefficient is a function of u, under some circumstances the
flux approximation can be derived using the nonlinear local BVP. In order to
have a generalization of the nonlinear flux scheme, it is essential to develop the
flux approximation for a generalized convective term as well as for nonlinear
diffusion.
108
Bibliography
[1] Adedoyin, A.A., Walters, D.K., and Bhushan, S. Investigation of

turbulence model and numerical scheme combinations for practical finite-
volume Large Eddy Simulations. Engineering Applications of Computa-
tional Fluid Mechanics 9 (2015), 324–342.
[2] Ascher, U.M., Ruuth, S.J., and Spiteri, R.J. Implicit-explicit

Runge-Kutta methods for time-dependent partial differential equations.
Applied Numerical Mathematics 25 (1997), 151–167.
[3] Bank, R.E., Rose, D.J., and Fichtner, W. Numerical methods

for semiconductor device simulation. SIAM Journal on Scientific and
Statistical Computing 4 (1983), 416–435.
[4] Bartuccelli, M.V., Doering, C.R., Gibbon, J.D., and Malham,

S.J.A. Length scales in solutions of the Navier-Stokes equations. Non-
linearity 6 (1993), 549–568.
[5] Batchelor, G.K. An Introduction to Fluid Dynamics. Cambridge

University Press, 2000.
[6] Bechmann, A. Large-Eddy Simulation of Atmospheric Flow over Com-

plex Terrain. PhD thesis, Technical University of Denmark, 2006.
[7] Black, F., and Scholes, M. The pricing of options and corporate
liabilities. Journal of Political Economy 81 (1973), 637–654.
[8] Brass, H., and Petras, K. Quadrature Theory: The Theory of Numer-
ical Integration on a Compact Interval, vol. 178 of Mathematical Surveys
and Monographs. American Mathematical Society, 2011.
[9] Brezzi, F., Marini, L.D., and Pietra, P. Two-dimensional expo-

nential fitting and applications to drift-diffusion models. SIAM Journal
on Numerical Analysis 26 (1989), 1342–1355.
109
BIBLIOGRAPHY
[10] Buelow, P.E.O., Venkateswaran, S., and Merkle, C.L. Stabil-

ity and convergence analysis of implicit upwind schemes. Computers &
Fluids 30 (2001), 961–988.
[11] Burgers, J.M. The Nonlinear Diffusion Equation: Asymptotic Solu-

tions and Statistical Problems. Reidel, 1974.
[12] Butcher, J.C. Coefficients for the study of Runge-Kutta integration

processes. Journal of the Australian Mathematical Society 3 (1963), 185–
201.
[13] Butcher, J.C. On the implementation of implicit Runge-Kutta meth-

ods. BIT Numerical Mathematics 16 (1976), 237–240.
[14] Butcher, J.C. Numerical Methods for Ordinary Differential Equations.

Wiley, 2008.
[15] Danet, C.P. The classical maximum principle. Some of its extensions
and applications. Annals of the Academy of Romanian Scientists Series
on Mathematics and its Applications 3 (2011), 273–299.
[16] Davidson, P.A. Turbulence: An Introduction for Scientists and Engin-

eers. Oxford University Press, 2004.
[17] Davis, P.J., and Rabinowitz, P. Methods of Numerical Integration.

Academic Press, 1984.
[18] de G. Allen, D.N., and Southwell, R.V. Relaxation methods ap-

plied to determine the motion, in two dimensions, of a viscous fluid past a
fixed cylinder. Quarterly Journal of Mechanics and Applied Mathematics
8 (1955), 129–145.
[19] Doedel, E.J. The construction of finite difference approximations to

ordinary differential equations. SIAM Journal on Numerical Analysis 15
(1978), 450–465.
[20] Doolan, E.P., Miller, J.J.H., and Schilders, W.H.A. Uniform

numerical methods for problems with initial and boundary layers. Journal
of Applied Mathematics and Mechanics 62 (1982), 280–281.
[21] El-Mistikawy, T.M., and Werle, M.J. Numerical method for

boundary layers with blowing - The exponential box scheme. AIAA
journal 16 (1978), 749–751.
[22] Eymard, R., Fuhrmann, J., and Gärtner, K. A finite volume

scheme for nonlinear parabolic equations derived from one-dimensional
local Dirichlet problems. Numerische Mathematik 102 (2006), 463–495.
110
BIBLIOGRAPHY
[23] Eymard, R., Gallouët, T., and Herbin, R. Finite Volume Methods,
vol. 7 of Handbook of Numerical Analysis. Elsevier, 2000.
[24] Farrell, P., and Linke, A. Uniform second order convergence of

a complete flux scheme on unstructured 1D grids for a singularly per-
turbed advection–diffusion equation and some multidimensional exten-
sions. Journal of Scientific Computing 72 (2017), 373–395.
[25] Frank, J., Hundsdorfer, W., and Verwer, J. On the stability of

implicit-explicit linear multistep methods. Applied Numerical Mathem-
atics 25 (1997), 193–205.
[26] Gartland, E.C. Uniform high-order difference schemes for a singularly

perturbed two point boundary value problem. Mathematics of Compu-
tation 48 (1987), 551–564.
[27] Gartland, E.C. Discrete weighted mean approximation of a model

convection-diffusion equation. SIAM Journal on Scientific and Statistical
Computing 3 (1982), 460–472.
[28] Geurts, B.J. An extended Scharfetter-Gummel scheme for high order

moment equations. COMPEL-The International Journal for Computa-
tion and Mathematics in Electrical and Electronic Engineering 10 (1991),
179–194.
[29] Ghia, U., Ghia, K.N., and Shin, C.T. High-Re solutions for incom-
pressible flow using the Navier-Stokes equations and a multigrid method.
Journal of Computational Physics 48 (1982), 387–411.
[30] Göçmen, T., van der Laan, P., Réthoré, P.E., Diaz, A.P.,
Larsen, G.C., and Ott, S. Wind turbine wake models developed
at the Technical University of Denmark: A review. Renewable and Sus-
tainable Energy Reviews 60 (2016), 752–769.
[31] Godunov, S.K., and Ryabenkii, V.S. Difference Schemes: An Intro-

duction to the Underlying Theory. Elsevier, 1987.
[32] Gogolides, E., and Sawin, H.H. Continuum modeling of radio-

frequency glow discharges. I. Theory and results for electropositive and
electronegative gases. Journal of Applied Physics 72 (1992), 3971–3987.
[33] Gresho, P.M., and Lee, R.L. Don’t suppress the wiggles - They’re
telling you something! Computers & Fluids 9 (1981), 223–253.
[34] Gresho, P.M., and Sani, R.L. On pressure boundary conditions for
the incompressible Navier-Stokes equations. International Journal for
Numerical Methods in Fluids 7 (1987), 1111–1145.
111
BIBLIOGRAPHY
[35] Gresho, P.M., and Sani, R.L. Incompressible Flow and the Finite
Element Method, vol. 2, Isothermal Laminar Flow. Wiley, 2000.
[36] Gupta, M.M., and Manohar, R.P. On the use of central difference
scheme for Navier-Stokes equations. International Journal for Numerical
Methods in Engineering 15 (1980), 557–573.
[37] Hairer, E., Lubich, C., and Roche, M. The Numerical Solution of
Differential-Algebraic Systems by Runge-Kutta Methods. Springer, 2006.
[38] Hamming, R.W. Numerical Methods for Scientists and Engineers.

Dover, 1986.
[39] Harlow, F.H., and Welch, J.E. Numerical calculation of time-

dependent viscous incompressible flow of fluid with free surface. Physics
of Fluids 8 (1965), 2182–2189.
[40] Harten, A. On a class of high resolution Total-Variation-Stable finite-

difference schemes. SIAM Journal on Numerical Analysis 21 (1984),
1–23.
[41] Hasager, C.B., Nygaard, N.G., Volker, P.J.H., Karagali, I.,

Andersen, S.J., and Badger, J. Wind farm wake: The 2016 Horns
Rev photo case. Energies 10 (2017).
[42] Heun, K. Neue Methoden zur approximativen Integration der Differen-

tialgleichungen einer unabhängigen Veränderlichen. Zeitschrift für Ange-
wandte Mathematik und Physik 45 (1900), 23–38.
[43] Huxley, L.G.H., and Crompton, R.W. The Diffusion and Drift of
Electrons in Gases. Wiley, 1974.
[44] Il’in, A.M. Differencing scheme for a differential equation with a small
parameter affecting the highest derivative. Mathematical Notes of the
Academy of Sciences of the USSR 6 (1969), 596–602.
[45] Jimenez, A., Crespo, A., Migoya, E., and Garcia, J. Large-Eddy
Simulation of spectral coherence in a wind turbine wake. Environmental
Research Letters 3 (2008).
[46] Keller, H.B. A new difference method for parabolic problems. In

Numerical Solution of Partial Differential Equations-II. 1971, pp. 327–
350.
[47] Kolmogorov, A.N. The local structure of turbulence in incompressible

viscous fluid for very large Reynolds numbers. Doklady Akademii Nauk
SSSR 30 (1941), 301–305.
112
BIBLIOGRAPHY
[48] Kumar, N., ten Thije Boonkkamp, J.H.M., and Koren, B. A sub-
cell discretization method for the convective terms in the incompressible
Navier-Stokes equations. In Lecture Notes in Computational Science and
Engineering (2016), 106, Springer, pp. 295–303.
[49] Kumar, N., ten Thije Boonkkamp, J.H.M., and Koren, B. A

new discretization method for the convective terms in the incompress-
ible Navier-Stokes equations. In Proceedings in Mathematics & Statistics
(2014), 77, Springer, pp. 363–371.
[50] Kutta, W. Beitrag zur Näherungweisen Integration totaler Differen-

tialgleichungen. Zeitschrift für Angewandte Mathematik und Physik 46
(1901), 435–453.
[51] Ladyzhenskaya, O.A. Solution “in the large” of the nonstationary

boundary value problem for the Navier-Stokes system with two space
variables. Communications on Pure and Applied Mathematics 12 (1959),
427–433.
[52] Ladyzhenskaya, O.A., and Silverman, R.A. The Mathematical The-

ory of Viscous Incompressible Flow. Gordon & Breach, 1969.
[53] Lambert, J.D. Computational Methods in Ordinary Differential Equa-

tions. Wiley, 1973.
[54] LeVeque, R.J. Finite Volume Methods for Hyperbolic Problems. Cam-
bridge University Press, 2002.
[55] Liu, L., van Dijk, J., ten Thije Boonkkamp, J.H.M., Mihailova,
D.B., and van der Mullen, J.J.A.M. The complete flux scheme-
Error analysis and application to plasma simulation. Journal of Compu-
tational and Applied Mathematics 250 (2013), 229–243.
[56] Mackenzie, D. An Engine, Not a Camera: How Financial Models Shape

Markets. MIT Press, 2006.
[57] Mahesh, K., Constantinescu, G., and Moin, P. A numerical

method for Large-Eddy Simulation in complex geometries. Journal of
Computational Physics 197 (2004), 215–240.
[58] Marx, Y.P. Time integration schemes for the unsteady incompressible
Navier-Stokes equations. Journal of Computational Physics 112 (1994),
182 – 209.
[59] Mattheij, R.M.M., Rienstra, S.W., and ten Thije Boonkkamp,

J.H.M. Partial Differential Equations: Modeling, Analysis, Computa-
tion. Monographs on Mathematical Modeling and Computation. SIAM,
2005.
113
BIBLIOGRAPHY
[60] Morinishi, Y. Conservative properties of finite difference schemes for

incompressible flow : 3rd report, numerical tests. Transactions of the
Japan Society of Mechanical Engineers Series B 62 (1996), 4106–4112.
[61] Morinishi, Y., Lund, T.S., Vasilyev, O.V., and Moin, P. Fully
conservative higher order finite difference schemes for incompressible flow.
[62] Morton, K.W., and Mayers, D. Numerical Solution of Partial Dif-

ferential Equations: An Introduction. Cambridge University Press, 2005.
[63] Patankar, S.V. Numerical Heat Transfer and Fluid Flow. Series in
Computational Methods in Mechanics and Thermal Sciences. Hemisphere
Pub. Corp., 1980.
[64] Polak, S.J., den Heijer, C., Schilders, W.H.A., and

Markowich, P. Semiconductor device modelling from the numerical
point of view. International Journal for Numerical Methods in Engineer-
ing 24 (1987), 763–838.
[65] Pope, S. Turbulent Flows. Cambridge University Press, 2000.
[66] Raithby, G.D. A critical evaluation of upstream differencing applied to

problems involving fluid flow. Computer Methods in Applied Mechanics
and Engineering 9 (1976), 75–103.
[67] Raizer, Y.P., and Allen, J.E. Gas Discharge Physics. Springer,
1997.
[68] Roache, P.J. Computational Fluid Dynamics. Hermosa, 1976.
[69] Roe, P.L. Approximate Riemann solvers, parameter vectors, and differ-
ence schemes. Journal of Computational Physics 43 (1981), 357–372.
[70] Roe, P.L. The use of the Riemann problem in finite difference schemes.
In Seventh International Conference on Numerical Methods in Fluid Dy-
namics: Proceedings of the Conference (1981), Springer, pp. 354–359.
[71] Roos, H.G., Stynes, M., and Tobiska, L. Robust Numerical Methods
for Singularly Perturbed Differential Equations: Convection-Diffusion-
Reaction and Flow Problems. Springer, 2008.
[72] Runge, C. Ueber die numerische Auflösung von Differentialgleichungen.

Mathematische Annalen 46 (1895), 167–178.
[73] Sagaut, P. Large Eddy Simulations for Incomressible Flows. Springer,

2002.
114
BIBLIOGRAPHY
[74] Sanderse, B. Energy-Conserving Discretization Methods for the Incom-

pressible Navier-Stokes Equations. PhD thesis, Eindhoven University of
Technology, 2013.
[75] Sanderse, B. Energy-conserving Runge-Kutta methods for the incom-

pressible Navier-Stokes equations. Journal of Computational Physics 233
(2013), 100–131.
[76] Sanderse, B., and Koren, B. Accuracy analysis of explicit Runge-

Kutta methods applied to the incompressible Navier-Stokes equations.
[77] Sanderse, B., van der Pijl, S.P., and Koren, B. Review of com-
putational fluid dynamics for wind turbine wake dynamics. Wind Energy
14 (2011), 799–819.
[78] Sanderse, B., Verstappen, R.W.C.P., and Koren, B. Boundary

treatment for fourth-order staggered mesh discretizations of the incom-
pressible Navier-Stokes equations. Journal of Computational Physics 257
(2014), 1472–1505.
[79] Sani, R.L., Shen, J., Pironneau, O., and Gresho, P.M. Pressure
boundary condition for the time-dependent incompressible Navier–Stokes
equations. International Journal for Numerical Methods in Fluids 50
(2006), 673–682.
[80] Schilders, W.H.A. Application of exponential fitting techniques to

semiconductor device simulation. Computational Methods for Boundary
and Interior Layers in Several Dimensions (1991), 175–202.
[81] Shigyo, N., and Niitsu, Y. Physical models for bipolar device simula-
tion and their effects on cutoff frequency. IEEE Transactions on Electron
Devices 40 (1993), 2087–2089.
[82] Spalding, D.B. A novel finite difference formulation for differential

expressions involving both first and second derivatives. International
Journal for Numerical Methods in Engineering 4 (1972), 551–559.
[83] Stock, H.W., and Haase, W. Determination of length scales in al-

gebraic turbulence models for Navier-Stokes methods. AIAA Journal 27
(1989), 5–14.
[84] Tafti, D.D. Comparison of some upwind-biased high-order formulations

with a second-order central-difference scheme for time integration of the
incompressible Navier-Stokes equations. Computers & Fluids 25 (1996),
647–665.
115
BIBLIOGRAPHY
[85] ten Thije Boonkkamp, J.H.M., Anthonissen, M.J.H., and

Kwant, R. A two-dimensional complete flux scheme in local flow adap-
ted coordinates. In Proceedings in Mathematics & Statistics (2017), 200,
Springer, pp. 437–445.
[86] ten Thije Boonkkamp, J.H.M., and Anthonissen, M.J.H. Exten-
sion of the complete flux scheme to time-dependent conservation laws. In
Numerical Mathematics and Advanced Applications 2009: Proceedings of
ENUMATH-2009 (2010), Springer, pp. 865–873.
[87] ten Thije Boonkkamp, J.H.M., and Anthonissen, M.J.H. The
finite volume-complete flux scheme for advection-diffusion-reaction equa-
tions. Journal of Scientific Computing 46 (2011), 47–70.
[88] ten Thije Boonkkamp, J.H.M., and Anthonissen, M.J.H. Nu-
merical dissipation and dispersion of the homogeneous and complete flux
schemes. In Proceedings in Mathematics & Statistics (2014), 77, Springer,
pp. 117–125.
[89] ten Thije Boonkkamp, J.H.M., and Schilders, W.H.A. An expo-
nential fitting scheme for the electrothermal device equations specifically
for the simulation of avalanche generation. COMPEL-The International
Journal for Computation and Mathematics in Electrical and Electronic
Engineering 12 (1993), 95–111.
[90] ten Thije Boonkkamp, J.H.M., van Dijk, J., Liu, L., and Peer-
enboom, K.S.C. Extension of the Complete Flux Scheme to systems of
conservation laws. Journal of Scientific Computing 53 (2012), 552–568.
[91] Thiart, G.D. Finite difference scheme for the numerical solution of fluid
flow and heat transfer problems on nonstaggered grids. Numerical Heat
Transfer, Part B: Fundamentals 17 (1990), 43–62.
[92] Thiart, G.D. Improved finite-difference scheme for the solution of
convection-diffusion problems with the SIMPLEN algorithm. Numerical
Heat Transfer, Part B: Fundamentals 18 (1990), 81–95.
[93] van Leer, B. On the relation between the upwind-differencing schemes
of Godunov, Engquist–Osher and Roe. In Upwind and High-Resolution
Schemes (1997), Springer, pp. 33–52.
[94] Veldman, A.E.P. “Missing” boundary conditions? Discretize first,
substitute next, and combine later. SIAM Journal on Scientific and
Statistical Computing 11 (1990), 82–91.
[95] Veldman, A.E.P. Computational Fluid Dynamics. Lecture notes in Ap-
plied Mathematics. Faculty of Mathematics and Natural Sciences, Uni-
versity of Groningen, 2013.
116
BIBLIOGRAPHY
[96] Verhulst, F. Methods and Applications of Singular Perturbations:

Boundary Layers and Multiple Timescale Dynamics, vol. 50 of Texts in
Applied Mathematics. Springer, 2005.
[97] Verstappen, R.W.C.P., and Veldman, A.E.P. Symmetry-

preserving discretization of turbulent flow. Journal of Computational
Physics 187 (2003), 343–368.
[98] Warming, R.F., and Hyett, B.J. The modified equation approach to
the stability and accuracy analysis of finite-difference methods. Journal
of Computational Physics 14 (1974), 159–179.
[99] Welch, J.E., Harlow, F.H., Shannon, J.P., and Daly, B.J. The
MAC method-A computing technique for solving viscous, incompress-
ible, transient fluid-flow problems involving free surfaces. Tech. rep., Los
Alamos Scientific Lab., 1965.
[100] Wesseling, P. Principles of Computational Fluid Dynamics. Springer,
2000.
[101] Wessels, F. Adaptive non-oscillatory complete flux schemes. Master’s
thesis, Eindhoven University of Technology, 2015.
[102] White, F. M. Viscous Fluid Flow. McGraw-Hill, 2006.

[103] Whitham, G. Linear and Nonlinear Waves. Wiley, 1999.
[104] Wilcox, D.C. Turbulence Modeling for CFD. DCW Industries, 1993.
[105] You, D., Mittal, R., Wang, M., and Moin, P. Analysis of stability
and accuracy of finite-difference schemes on a skewed mesh. Journal of
Computational Physics 213 (2006), 184–204.
117
Summary
Flux approximation schemes for flow problems using local boundary

value problems
In this thesis, we develop discretization methods for nonlinear flow problems.

This is inspired by the need for even more accurate and efficient numerical
methods for simulating the flow of incompressible viscous fluids governed by
the Navier-Stokes equations. The approach taken is to construct a flux approx-
imation scheme, where the numerical flux is computed from a local boundary
value problem (BVP). The use of local BVPs for the approximation of the
fluxes allows us to include several effects on the flux, for instance, the pressure
gradient and the gradient of the transverse flux, through source terms. For un-
steady flow problems, the acceleration can also be regarded as a source term.
The local BVP methods are shown to be stable and accurate, even on coarse
meshes, and do not introduce artificial diffusion.
In the first part of this thesis, we discuss the discretization of the convect-
ive terms in the incompressible Navier-Stokes equations by iteratively solving
linearized local two-point BVPs to compute the cell-face velocity components.
The local BVP in this case can be formulated using the momentum equation
and taking the pressure gradient and the gradient of the transverse flux as the
source terms. In the absence of source terms we have a homogeneous local
BVP, the interface velocity is then given by a weighted sum of the first-order
upwind and the second-order central scheme. The weight function for the ho-
mogeneous approximation depends directly on the mesh Péclet number, that
is the local ratio of the convective and viscous forces. Consequently, the ho-
mogeneous approximation is at most second-order convergent for finer meshes
and first-order on coarse meshes. The accuracy of the approximation can be
improved by the inclusion of the source terms, i.e., the pressure gradient and
the gradient of the transverse flux. Since the underlying local BVP is a linear
problem, the interface velocity can be expressed as the sum of a homogeneous
and an inhomogeneous part, with the latter depending on the source terms.
The inclusion of the source terms provides a more accurate approximation for
the interface velocities, with the inhomogeneous approximation being uniformly
118
Summary
second-order convergent. Moreover, the inhomogeneous approximation shows

significantly lower numerical diffusion than the homogeneous approximation.
The second part of this thesis addresses the use of the complete-flux schemes
for nonlinear flow problems. In the complete-flux scheme the flux, i.e., the sum
of the convective and the viscous flux, is approximated using local BVPs. To
simplify the derivation, the nonlinear convective terms are linearized using cell-
face velocities, that are computed from local BVPs. As before, the linearity of
the model BVP allows us to express the flux as the sum of a homogeneous part
and an inhomogeneous part, with the former being a weighted sum of the first-
order upwind and the second-order central scheme and the latter depending on
the source terms in the BVP. The inclusion of the source terms provides a better
approximation, with the inhomogeneous scheme being uniformly second-order
convergent.
The third part of the thesis discusses the flux approximation for nonlinear
BVPs. As a result of the nonlinearity of the BVP, the flux approximation can
not be expressed as the sum of a homogeneous and an inhomogeneous part.
Moreover, the nonlinearity poses another challenge, that is to have a consistent
flux function that reduces to the correct limit flux, for vanishing viscosity. We
address both problems and show the existence of a consistent nonlinear flux
approximation for the viscous Burgers equation. The nonlinear approximation
is shown to be more accurate than the linear schemes.
The ideas put forward in this thesis provide an accurate and robust discret-
ization method for nonlinear flow problems. The use of local BVPs gives an
approximation that depends directly on the mesh Péclet number and includes
the effects of the source terms. Using nonlinear local BVPs gives us compact
schemes (three-point in each spatial direction) that are second-order accurate
and free of spurious oscillations.
119
Curriculum Vitae
Nikhil Kumar was born on 21 July 1990 in Jaipur, India. He finished his
schooling from Gyan Bharti School, New Delhi, India. He subsequently stud-
ied Mathematics at the Indian Institute of Science Education and Research
(IISER), Mohali, India and obtained a master’s degree in 2013, with a master
thesis entitled Computational Aspects of Representation Theory.
In September 2013, Nikhil started his PhD project at Eindhoven University
of Technology, under the supervision of dr.ir. Jan ten Thije Boonkkamp and
prof.dr.ir. Barry Koren. The results obtained during the PhD project are
presented in this dissertation. The PhD position was funded under the Shell-
NWO “Computational sciences for energy research (CSER)” initiative.
121
List of publications
Journal Publications:
1. N. Kumar, J.H.M. ten Thije Boonkkamp and B. Koren, Approximation
of the convective flux in the incompressible Navier-Stokes equations using
local boundary-value problems, Journal of Computational and Applied
Mathematics (Submitted, 2017).
2. N. Kumar, J.H.M. ten Thije Boonkkamp and B. Koren, Nonlinear flux
approximation scheme for the viscous Burgers equation, (in preparation).
Refereed proceedings:
1. N. Kumar, J.H.M. ten Thije Boonkkamp, B. Koren and A. Linke, A
nonlinear flux approximation scheme for the viscous Burgers equation,
Proceedings in Mathematics & Statistics (200): 457−465, Springer, 2017.
2. N. Kumar, J.H.M. ten Thije Boonkkamp and B. Koren, Flux approxim-
ation scheme for the incompressible Navier-Stokes equations using local
boundary value problems, Lecture Notes in Computational Science and
Engineering (112): 43 − 51, Springer, 2016.
3. N. Kumar, J.H.M. ten Thije Boonkkamp and B. Koren, A sub-cell dis-

cretization for the convective terms in the incompressible Navier-Stokes
equations, Lecture notes in Computational Science and Engineering (106):
295 − 303, Springer, 2015.
4. N. Kumar, J.H.M. ten Thije Boonkkamp and B. Koren, A new discret-
ization method for the convective terms in the incompressible Navier-
Stokes equations, Proceedings in Mathematics & Ststistics (77): 363−371,
Springer, 2014.
123
Acknowledgments
I consider myself very fortunate to be surrounded by some incredible people,

who kept me motivated all along the journey of the PhD.
First and foremost, I am grateful to my supervisors, Jan and Barry, who
were always there to show me the path, whenever I felt lost. Both of you have
set a very high bar for my future supervisors/bosses.
Jan, it was a pleasure working with you and I am grateful for your con-
tribution to my research. Your guidance over the past four years has deeply
influenced me both academically and personally. It was always nice to travel
with you for conferences and have some insightful dinner discussions.
Barry, your foresight was very helpful in linking the problems and drawing
a sketch of the thesis long before I set upon the task of writing it. Thanks for
reading and editing all the drafts, I am sure I will miss your colorful corrections.
I would like to thank the personnel involved in the Shell-NWO Computa-
tional sciences for energy research (CSER) PhD program for providing me an
opportunity to be a part of this combined initiative between Dutch and Indian
research communities. Specially, I am thankful to Paul Mak (Shell Projects &
Technology), Joost Weber (NWO-I) and Maria Teuwissen (NWO-I) for helping
me throughout the various stages of the CSER PhD program.
I would also like to thank my committee members: prof. Kees Vuik, prof.
Arthur Veldman, prof. Harald van Brummelen and dr. Benjamin Sanderse for
finding time to read my thesis and providing their valuable feedback. I am also
thankful to Dr. Alexander Linke for his inputs in the later part of my research,
working with you was a pleasant experience.
CASA (Centre for Analysis, Scientific computing and Applications) always
felt like a home away from home and I will always cherish the memories of
my time spent at CASA. Enna, you always helped me with all nitty-gritties of
the administrative work, which would have been a nightmare without you. I
would like to thank all my colleagues, as well as the past CASA members for
the jollity at the lunch tables and around the coffee machines. Special thanks
to my office mates Bart, Rien and past office mates Corien, René and Frank
for the numerous interesting ‘small talks’ (specially the ones after beers) and
also for translating all the documents into Dutch.
125
I am also thankful to my friends Abhineet, Amar, Abhijay, Anwesha, Tarun,
Koondi, Pranav, Stephen, Sai, Marco, Fritzzie, Carina, Jeroen, Wen, Nenad,
Manu, Deepesh and Sumit for making a stay in Eindhoven a memorable one.
In the end, I am forever grateful to my parents, for their love and support,
all throughout my life. They were always there to give me the push whenever,
it got a bit difficult. Priya and Pravesh (my sister and brother-in-law) both
of you were always there to give the right advice and I can see that trait has
been very well passed to Saisha (my niece) who has already started giving me
advice(s) and she’s just 2!!
Finally, Netherlands thank you for the fantastic bike paths and the nice
cafés along the bike path, I am sure I will miss them very much.
126

Flux Approximation Schemes For Flow Problems Using Local Boundary Value Problems

Uploaded by

Copyright:

Available Formats

Flux Approximation Schemes For Flow Problems Using Local Boundary Value Problems

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Flux Approximation Schemes For Flow Problems Using Local Boundary Value Problems

Uploaded by

Copyright:

Available Formats

Flux approximation schemes for flow problems using

local boundary value problems

Please check the document version of this publication:

Citation for published version (APA):

Download date: 01. Jan. 2019

A catalogue record is available from the Eindhoven University of Technology

Copyright © 2017 by N. Kumar.

ter verkrijging van de graad van doctor aan de

geboren te Jaipur, India

voorzitter: prof.dr. M.A. Peletier

Het onderzoek dat in dit proefschrift wordt beschreven is uitgevoerd

4 Convective discretization of the Navier-Stokes equations 37

6 Complete-flux scheme for the Navier-Stokes equations 77

7 Nonlinear flux approximation 87

Curriculum Vitae 121

List of publications 123

The conservation laws governing fluid flow involve a combination of convective

• Drift-diffusion model. The drift-diffusion model is used for describing

• Black-Scholes model. The Black-Scholes model, first published by Fisher

for the incompressible Navier-Stokes equations, given by:

1.2 Spatial discretization

• have a three-point coupling, corresponding to each spatial direction,

• have as little numerical dissipation as possible,

• be uniformly second-order accurate,

1.3 Main results

Incompressible Navier-Stokes equations. The nonlinear convective term

accuracy of the homogeneous approximation, resulting in a uniformly second-

Viscous Burgers equation. We derive the integral representations using

1.4 Outline of the thesis

In this chapter, we present an overview of the spatial and temporal discretiz-

2.1 Convection-diffusion equation

φi−2 φi+1 φi+2

Figure 2.2: Cell-centered finite-volume discretization of the 1-D convection-

Discrete maximum principle. For robust numerical methods, it is often

Proof. See chapter 4, [100].

2.1.1 Time integration

where Un ≈ U(tn ) and ki is defined as

c1 a11 a12 ··· a1s

As the order of the method increases, the number of algebraic conditions on

mainly because the higher-order of accuracy allows us to ignore the discretiza-

(I + ∆tθA)Φn+1 = I − ∆t(1 − θ)A Φnh + ∆t θSn+1 + (1 − θ)Snh +

2.2 Incompressible Navier-Stokes equations

(M u)i,j = ūi+1/2,j − ūi−1/2,j + ṽi,j+1/2 − ṽi,j−1/2 , (2.19)

where ūi+1/2,j and ṽi,j+1/2 represent the face-averaged velocity components

ūi+1/2,j = ∆y ui+1/2,j , ṽi,j+1/2 = ∆x vi,j+1/2 .

(Du (u))i+1/2,j = δx u i+1,j − δx u i,j + δg

Gu (p)i+1/2,j = p̄i+1,j − p̄i,j , Gv (p)i,j+1/2 = p̃i,j+1 − p̃i,j . (2.22)

The discretization scheme can be extended to achieve higher-order accur-

Lp = M F(u, t) − ṙ1 (t), (2.23)

where L = M G is the discrete Laplacian operator and

F(u, t) = −N(u) + Du(t) + r2 (t). (2.24)

2.2.1 Time integration

0 =g(u, t), (2.25a)

g(u, t) =Mu − r1 (t),

with F(u, t) as defined in equation (2.24). In an s-stage Runge-Kutta method

In fact, Φi is a Lagrange multiplier to make Ui divergence-free, with each

The finite-volume discretization of conservation laws requires fluxes to be ap-

3.1 Convection-diffusion equation

L[φ] := aφx − νφxx = s, x ∈ (0, L), (3.1)

F(u, t) = −N(u) + Du(t) + r2 (t). (2.24)

ω(κ) = aκ + κ3 6= const.

(u2 − ux )x = −px − (uv − uy )y , = 1/Re, (4.5a)

V uy − uyy = 0, yj < y < yj+1 , x = xi+1/2 , (4.22a)

U vx − vxx = 0, xi < x < xi+1 , y = yj+1/2 , (4.23a)